Original research, practical explanations and interactive experiments about artificial intelligence, AI agents, machine autonomy and their implications for people and institutions.
AI used to answer individual questions. It can increasingly carry out whole workflows. What changed is not just the quality of the answer: AI can now access knowledge, use software tools, plan a sequence of steps and work towards an outcome.
Why it matters: Businesses should stop asking only “How can staff use AI?” and start asking “Which parts of our workflows can AI reliably complete?”
The important distinction: More capability does not automatically justify more independence. Permissions and oversight still matter.
At each stage, more of the steps between an instruction and a completed task can be performed by AI.
In our guide to the six SAE levels of self-driving cars, each level showed how more of the driving task could pass from human to machine. AI presents a similar question for knowledge work: how much of the task can software now carry through itself? Unlike SAE driving levels, the six AI stages below are a WiBrief teaching framework, not an official standard.
The six stages: from prompt to completed workflow
01 — Answer
A single prompt produces a response. Human work: ask, assess, copy, act and initiate the next prompt.
Client proposal example: “Write an introduction to a client proposal.”
02 — Converse
The assistant carries context across a discussion, clarifies requirements and improves an answer over several turns. Human work: steer the conversation and execute the task.
Client proposal example: “Ask me about the client, then refine the introduction.”
03 — Retrieve
The assistant searches connected sources to ground its work in documents, facts and precedents. Human work: verify evidence and apply the result.
Client proposal example: “Find relevant previous fee proposals and use them as references.”
04 — Use tools
The system can call approved software functions, APIs and integrations, potentially via MCP. It can take specific actions rather than only recommend them. Human work: specify tasks, grant access and review consequential actions.
Client proposal example: “Retrieve approved records from SharePoint and create a working draft.”
05 — Plan and coordinate
Given a goal, an agent can select and sequence multiple tools, inspect results and revise its plan when something fails. Human work: define objectives, constraints and exceptions.
Client proposal example: “Search precedents, draft the proposal, check pricing rules and flag missing details.”
06 — Complete workflows
Within pre-agreed scope, an agent or collection of agents can run a workflow from request to defined outcome, including monitoring progress and escalating exceptions. Human work: set policy, approve high-impact decisions and supervise outcomes.
Client proposal example: “Prepare the proposal package, route it for approval and finish the internal workflow.”
The change in one picture
In the early stages: Human → AI answer → Human completes the work.
In the middle: Human → AI uses knowledge and tools → Human reviews progress.
At the advanced stages: Human sets goal and boundaries → AI plans and executes steps → Human supervises outcomes and exceptions.
This is a functional progression, not a strict historical timeline: conversation, retrieval and tool-use capabilities overlap, and autonomy depends on the system’s design and permissions.
What actually made this possible?
Model capability improved reasoning and instruction-following; longer context and retrieval helped systems work with relevant information; tool calling and MCP-style connections opened access to software; and agent architectures added loops for planning, acting, observing and correcting. None of these alone guarantees reliable autonomy.
What UK research says
The UK government’s AI Insights: Agentic AI, updated in August 2026, explains how language-model capabilities combine with agent architectures to enable systems to select and coordinate actions. The March 2026 consumer research frames the shift as moving from tools that support user decisions towards agents potentially entrusted with outcomes.
The UK’s AI Security Institute examined more than 1,000 publicly available MCP servers and reported increasing levels of autonomy among newly published finance-related tool interfaces across part of 2025. That is evidence about tooling and permissions, not proof that all deployed agents perform reliably.
The National Cyber Security Centre advised in May 2026 that organisations should start with tightly bounded, low-risk uses, maintain visibility and retain clear human responsibility. Its 20 October 2026 research problem book update asks how semi-autonomous cyber systems can be evaluated and assured.
Take one process with a clear input and outcome. Count how many steps currently require a human. Identify which steps AI can perform safely today, which need approval, and which are not ready for delegation. Measure completed outcomes, not impressive demonstrations.
The destination is not a smarter chatbot. It is a dependable system capable of completing useful work.
Editorial note: This six-stage framework is an explanatory model developed for WiBrief. It should not be confused with a regulatory classification or a validated capability benchmark. Sources checked 22 October 2026.
AI is becoming increasingly capable at producing answers. Yet organisations still struggle with a more consequential question: how do we turn those answers into decisions we can trust?
The missing layer between intelligence and impact
Consider a team adopting generative AI. One person uses it to summarise meetings; another drafts proposals; a third analyses documents. Each task becomes faster. But nothing necessarily connects the meeting to the proposal, the proposal to the decision, or the decision to what the organisation learns next.
The limitation is not simply the model. It is the absence of a deliberate system connecting evidence, expertise, action and feedback.
The decision intelligence loop
CAPTURE → STRUCTURE → EVALUATE → DECIDE → LEARN
Evidence becomes structured options. Human judgement determines action. Outcomes improve the next decision.
A practical architecture for better decisions
Capture: Collect the relevant facts, assumptions and sources at the point where work happens.
Structure: Translate scattered observations into comparable opportunities, risks and options.
Evaluate: Assess each option against explicit criteria such as impact, effort, confidence and ethical consequences.
Decide: Make responsibility visible. AI may inform a decision; it should not silently inherit authority to make it.
Learn: Compare what was expected with what actually happened, and update the process.
This is a decision system: a repeatable pathway from information to accountable action. AI can accelerate every stage, but the pathway itself is a design choice.
Why prompts alone are insufficient
A prompt is an instruction within a moment. A system preserves continuity across moments. It can carry provenance, constraints, ownership, escalation routes and feedback. This distinction matters wherever outcomes must be explained, challenged or improved.
For example, when designing a client proposal, the important question is not merely whether AI can draft persuasive copy. It is whether the team can trace which client needs, prior examples, pricing assumptions and human approvals informed the final proposal.
Three tests for responsible design
Traceability: Can someone reconstruct the evidence behind a recommendation?
Contestability: Can a knowledgeable person question, reject or override it?
Learning: Does the system improve from real outcomes rather than from confidence alone?
These are practical tests, not guarantees. A neatly structured workflow can still encode poor assumptions or institutional bias. Good governance therefore requires periodic examination of who benefits, who bears risk and whose perspectives are missing.
Build one small decision loop this week
Choose one recurring decision in your work. Write down its inputs, the criteria used to assess options, the person accountable, and the outcome you will check later. Only then ask where AI could remove friction. You may discover that the highest-value improvement is not a more sophisticated model but a clearer hand-off or a more consistent evaluation method.
The future of effective AI adoption may be determined less by how fluently machines respond and more by how deliberately we design the human systems around them.
The Digital Revolution. When Apple introduced the iPhone, it wasn’t just about making calls. It changed how we work, learn, shop, navigate, and even how we relax. It made the internet personal.
Now, we’re on the edge of another leap forward—this time with artificial intelligence. But instead of fitting into one device, AI has the potential to fit around you. The real magic? Custom GPTs. These are AI assistants tailored to your preferences, your data, your goals.
Much like the shift from landlines to smartphones, we’re shifting from generic AI to personalised, reliable assistants that can enhance productivity in ways that feel surprisingly human.
2. What Is a Custom GPT Persona?
A GPT persona is a customised version of a general AI assistant like ChatGPT. Instead of just answering questions with whatever information it can recall, a persona is:
Aligned with your tone, style, or domain
Pre configured with your custom instructions, documents and preferences
Built to follow certain rules or avoid specific mistakes
Think of it like your own digital co-worker who already knows how you think, what you care about, and how you like things done each time you use it.
Just like Netflix recommends based on your viewing habits, a GPT persona can:
Help you prepare presentations from your notes
Answer questions from your uploaded course textbook
Provide health insights based on your fitness data
Extract the “important bits” from long contracts
Draft emails, lesson plans, or research summaries the way you would
You don’t need to be technical to shape AI. You just need to be clear about what you want it to do—and where you still want to stay in control. Because the real risk isn’t AI doing too little. It’s AI quietly making decisions for you.
🖼️ Where to Start: Build Your Own GPT Assistant Visit https://chat.openai.com/gpts and click “Explore GPTs” to browse examples or “Create a GPT” to build your own in minutes—no code required.
Use the guided builder to define its tone, purpose, and upload key documents or links. You can even give it a name and a profile picture!
3. What Powers This? Meet RAG (Without the Jargon)
Behind the scenes, many custom GPTs are supercharged with something called Retrieval-Augmented Generation (RAG). Don’t let the name scare you.
All it means is this:
You can upload your own content (a PDF, notes, slide decks)
Or, link it to web pages or online sources you trust
The AI then uses that content, not just what it was pre-trained on
So instead of generic or outdated info, the assistant pulls exactly what it needs from your sources.
🔍 Real-Life Uses of RAG + Personas
These combinations unlock powerful everyday tools:
🏋️ Fitness Coach GPT: Upload your health plan or workout logs and talk to a coach that helps you stay on track and explains your stats.
📚 Study Buddy GPT: Upload your course curriculum and textbooks—ask questions, quiz yourself, and get summaries tailored to your syllabus.
📅 Presentation Assistant GPT: Drop in your research or agenda and get a draft presentation, bullet points, or speaker notes.
🔒 Contract Navigator GPT: Upload a bunch of boilerplate contracts and get the unusual or noteworthy provisions flagged for you.
📜 Personal Researcher GPT: Paste in reports, links, or documents, and it will help summarise, compare, or explain them in your preferred format.
These are low-barrier, high-impact applications. No coding. No jargon. Just results that make sense for you.
But remember: AI sees patterns. It doesn’t feel instinct. It might catch six common risks in a document—but miss the one that’s subtly phrased, buried in boilerplate, or just feels off. That’s why your judgment still matters.
4. Multiple Personas, Working as a Team
Now imagine this: not just one assistant, but a team of them working behind the scenes.
One gathers relevant info
Another analyses and drafts
A third reviews and refines
This is already happening across industries. And it’s how professionals are getting reliable, repeatable results from AI—the kind you can trust.
But it’s just as relevant for students, retirees, freelancers, or anyone managing tasks that require judgment, organisation, or insight.
And just like a GPS uses traffic data to help you avoid delays, your personas can adapt in real time, making your work smoother, smarter, and more aligned with how you think.
5. Don’t Forget the Human Touch
AI is fast. Sometimes too fast.
It’s tempting to accept whatever it gives you—especially when it sounds confident. But speed doesn’t equal accuracy.
Think of AI like a smart intern: helpful, yes—but you still need to review its work. A model might flag six obvious red flags in a contract… but miss the one that’s subtle, strange, or completely unexpected. That’s where your judgment matters most.
So as you delegate to AI, don’t outsource your thinking. Build in moments to pause and ask: “Is this right?” “Is anything missing?” “Would I say it this way?”
That’s not friction. That’s good decision-making.
AI should amplify your instincts—not override them.
6. The Takeaway
Custom GPTs are more than cool tech. They’re a way to bring clarity, speed, and insight to your everyday work and life.
We no longer need to adapt to AI. Now, AI can adapt to us.
And the real productivity boost isn’t just from automation. It’s from intelligent delegation—handing off the repetitive parts of your work so you can focus on what matters: judgment, creativity, strategy.
Whether you’re a teacher prepping lessons, a student revising for finals, or a founder writing investor updates, the question is the same:
“Is this AI helping me think better—or just think less?”
The best custom GPTs don’t just save time. They sharpen your focus, elevate your thinking, and help you do more of what only you can do.
If the iPhone put the world in your pocket, custom GPTs put you at the centre of your digital universe.
Continue exploring
Custom AI assistants are one step along a broader spectrum of machine autonomy. See how the stages of delegation fit together in the Autonomy and AI research hub, and investigate decision authority in the interactive Autonomy Atlas.
As GenAI digital assistants—such as Microsoft Copilot, vendor-specific versions of ChatGPT, and internal firm models—embed themselves into legal workflows, the question is no longer “Should I use AI?” It’s: How do we design it to enhance—not override—judgment? The future of professional services depends not just on what AI can do, but on how we choose to think.
⚖️ From Oversight to Ownership
In professional services, our value has never been just in what we know—but in how we frame, filter, and challenge complexity. The rise of GenAI digital assistants doesn’t erase that—it raises the stakes.
Because even if:
AI proposes a new content and structure based on your parameters,
Ranks risk based on frameworks you define,
Surfaces insights drawn from analogue experience…
The question remains:
Are you still making the key decisions—or just reviewing what the system has already prioritised for you?
This is what I call judgment drift—not losing control outright, but gradually letting automated systems decide what matters most in professional decisions—without continuous, active human evaluation for real-world relevance.
🔍 A Real-World Example
AI can quickly scan hundreds of pages and flag six familiar risks—termination clauses, liability caps, change-of-control provisions. But what about the seventh? The one that doesn’t follow the usual pattern. The clause phrased just differently enough, buried in boilerplate, or introduced in passing by a counterparty. That’s the one the model might miss—and the one a sharp human spots, because it feels off.
In high-stakes negotiations, using GenAI to draft too early can give the illusion of a well-framed position—when in reality, it’s just a statistically likely one. That false confidence can lock you into assumptions on liability, pricing, or responsibility—before you’ve aligned the output with your client’s goals, strategy, or leverage.
The result? Time saved, but oversight lost.
That’s why the goal isn’t automation for its own sake. It’s intelligent delegation—with human judgment guiding what to review,when to slow down, and where to challenge the default.
✨ Story Moment: Friday Evening. 900 Pages. No Panic.
You’ve lived this.
It’s late Friday. A critical deliverable just landed—a 900-page document, a dataset, a transaction report. It needs to be reviewed, structured, and summarised by Monday.
Old world? Two team members brace for a lost weekend. New world? Your organisational GenAI assistant processes the file, extracts what matters, flags inconsistencies, and delivers a structured first-draft summary—in your specified standard format—before you even log off.
You’re not replacing expertise. You’re reclaiming it—for strategy, not survival.
🛠️ What Should Professional Services Do Next?
1. Shift from Passive Tool Use to Active Role Design
Don’t just use AI—author it. Define the assistant’s role, purpose, tone, steps to perform and boundaries. Gen AI tools perform better when shaped by clear context and intentional structure—not guesswork.
2. Preserve Cognitive Friction
Build in moments to pause, question, and challenge what the Gen AI system returns. Good judgment doesn’t come from seamless flow—but from thoughtful interruption.
3. Create Frameworks Worth Scaling
Treat every effective Gen AI setup as a blueprint—and build from there. Curate prompts that work, regardless of the platform. Adapt and improve them.Document the logic. Share the method.Build systems others can trust—and scale.
🧭 Why This Matters
GenAI is no longer speculative—it’s operational.
The professionals who will lead this shift aren’t the ones who use AI fastest. They’re the ones who ask:
What’s being missed?
Who is really in control?
Does this support or replace my judgment?
and most importantly:
Is it amplifying your expertise—or quietly eroding the very judgment your clients rely on?
Top 10 Predictions Shaping Legal, AI & Human Agency (2025–2028)
TL;DR
As GenAI remakes the legal and professional landscape, the competitive edge is shifting from mere automation to intentional, governed innovation. The winners will be those who embed robust oversight, transparent validation, and dynamic governance frameworks into the very fabric of their work—reclaiming human agency and public trust.
1. Human Oversight as a Competitive Service
By 2028, “human oversight” will be as marketable as legal expertise or technological prowess. Clients will increasingly demand visible, auditable proof of where, when, and how critical human judgment is applied—spanning from contracts to compliance reviews. “Oversight-as-a-Service” will become a selling point, with governance models showcased in RFPs and public disclosures.
2. Governance Becomes the DNA of Innovation
No longer a compliance afterthought, governance moves to centre stage—shaping not just what gets built, but how it’s built, tested, and scaled. Law firms and legal tech teams will institutionalise dynamic governance: real-time feedback loops, champion validation, transparent escalation protocols, and proactive audit trails. This enables both safe experimentation and trusted adoption.
3. Strategic Cognitive Offloading Defines Talent Value
The most valued professionals will not just use AI, but demonstrate discernment—knowing when to offload to machines and when to intervene as a human. Expect advancement frameworks to explicitly reward this judgment, codified through governance structures that track, review, and continuously recalibrate the human-AI handoff.
4. Kitemarked Outputs and Visible Validation
“Human Oversight Certified” seals and similar kitemarks will become compliance baselines in legal, audit, and high-risk advisory work. These markers, only awarded through transparent, traceable validation frameworks, will be a sign of trust—not mere decoration.
5. The Return of ‘Slow Work’—By Design, Not Default
To counteract the deskilling risks of automation, leading firms will set up deliberate slow lanes: protected time and space for deep legal analysis, mentorship, and creative reasoning. These slow work zones will be governed for quality—tracked and measured, not left to chance.
6. Outcome-Based Pricing—Driven by Transparent Process
AI-enabled productivity will accelerate the move away from the billable hour. Value-based pricing will require transparency about how work is delivered—AI, human, or hybrid—and governance documentation that clients can review.
7. Living Sensemaking Hubs Replace Static Knowledge Bases
Platforms like WiBrief will move from static content to living foresight engines, integrating scenario modeling, prediction scoring, and transparent governance of how insights are generated and validated. This continual auditability will be a marker of platform trust.
8. Public Self-Audit and Transparent Reflection
Annual self-audits, open prediction scoring, and reflective learning cycles will become professional norms. Firms will publicise both successes and misses—anchoring trust in governed humility, not just glossy marketing.
9. Bias Exposure Drives Radical Transparency in Legal Advice
AI will illuminate unseen biases in both algorithmic and human outputs. The most trusted organisations will not just acknowledge this risk, but embed bias detection and remediation in their governance models, making correction and learning as visible as outcomes.
10. Leadership Redefined: From Command to Coherence
The new leaders will not be those who control more AI, but those who engineer coherence—integrating ethical frameworks, oversight rituals, and transparent governance into every layer of decision-making. Strategic enablement sprints, champion-driven reviews, and participatory governance will distinguish cultures that thrive.
The Golden Thread:
In every prediction above, robust governance is not a box to tick—it is the architecture of trust. The firms and professionals who make their frameworks visible, participatory, and continuously evolving will shape the next decade—not just by what they predict, but by how they keep themselves accountable to those predictions.
🧭 Ready to benchmark your own approach? Explore the WiBrief GenAI Governance Compass — a practical, self-assessment framework to help leaders, teams, and organisations navigate responsible GenAI adoption and innovation.
Big Question for 2025–2028:
Are you just adopting GenAI—or are you building a living system where innovation, judgment, and governance rise together?
Benchmark your readiness. Navigate responsible innovation.
As organisations race to adopt generative AI, the difference between thriving and stumbling isn’t just technology—it’s governance. This self-check tool enables leaders, teams, and innovators to evaluate current practices, identify gaps, and set a principled course for trustworthy, high-impact AI adoption.
Use this Compass to align your strategy, processes, and culture with the standards that will define tomorrow’s most trusted organisations.
1. Clear Purpose and Role Definition
Are the objectives, boundaries, and expected outcomes of your project explicitly documented?
Are all key roles (owners, validators, reviewers) clearly assigned and communicated?
2. Structured Idea Capture
Is there an open, documented process for collecting and recording new use cases, challenges, or improvements from all stakeholders?
Are proposed ideas logged transparently in a central repository?
3. Responsible Prototyping and Risk Awareness
Are new ideas moved quickly from concept to draft—while identifying risks, data requirements, and oversight needs at each step?
Is there a framework for assessing risk and value before scaling or deployment?
4. Independent Validation
Does each major output undergo review by a qualified expert or peer, independent from its creator?
Are feedback cycles documented, with a clear record of decisions and improvements?
5. Transparent Feedback and Iteration
Is regular review and retrospective discussion built into the process, ensuring that lessons and concerns are surfaced and addressed?
Are critical issues or deviations flagged early and managed transparently?
6. Auditable Oversight
Are oversight and validation steps clearly documented and visible to all relevant parties—including clients, where appropriate?
Is there traceability of decision points and interventions throughout the lifecycle?
7. Continuous Self-Audit and Reflection
Are outcomes periodically assessed and reported, both for successes and learning opportunities?
Is there a culture of open reflection, transparency, and accountability around both achievements and setbacks?
Self-Assessment (Rate 1–5 for Each Area)
1 = Not evident
3 = Partially in place
5 = Fully embedded
Scoring Guide:
30–35:Governance Pioneer – Strong, consistent frameworks support trusted, resilient innovation.
22–29:Progressing – Key governance elements present; targeted improvements will enhance trust.
15–21:Developing – Governance is emerging but uneven; focused attention required.
Below 15:At Risk – Gaps may expose the organisation to avoidable risks and erode stakeholder trust.
Reflection & Next Steps
Which governance area is your greatest strength—and which needs the most urgent attention?
What practical action can you take this quarter to strengthen your governance baseline?
Consider sharing your results (anonymously or otherwise) to contribute to a wider culture of transparency and improvement.
🚀 Artificial Intelligence (AI) is not just changing the game—it’s transforming the entire playing field. 🌍 From healthcare breakthroughs to smarter financial systems, AI is driving innovation and efficiency at an unprecedented pace. But with this rapid advancement comes a crucial responsibility: ensuring that AI is used ethically, safely, and within the bounds of robust regulatory frameworks.
This page is your comprehensive guide to some of the most powerful, publicly accessible AI tools. Here, you’ll uncover the cutting-edge functionalities of these tools, explore their potential risks, and understand the regulatory frameworks that will help you navigate this complex and dynamic field. Whether you’re a developer, researcher, or AI enthusiast, this guide is designed to keep you informed and ahead of the curve.
🔍 Category
🛠️ AI Tool/Platform
💡 Description
🌟 Why It Matters
📜 Regulatory Considerations
🔗 Learn More
🔓 Open-Source AI Models
Hugging Face
A hub for community-driven models and datasets now featuring integrations with GPT-4.5 and advanced multimodal models.
Democratizes AI innovation and enables collaborative development across industries.
Supports ethical open-source licensing and aligns with emerging global transparency standards.
A useful AI tool is not automatically a reliable decision-maker. Explore how assistants are configured, how to evaluate their consequences and why clear governance matters.
📂 Remember the days spent sifting through endless files or scouring outdated databases for that one crucial piece of information? Whether you’re a 💼 lawyer, a 🔬 researcher, a 📈 marketer, or any professional who relies on knowledge, we’ve all been there. 👋 Say goodbye to those frustratingly slow and inefficient workflows.
🚀 The game-changer is here: Retrieval-Augmented Generation (RAG). This cutting-edge AI technology is set to revolutionise how we work across various fields, giving us instant access to up-to-date and relevant knowledge right at our fingertips. It’s like strapping a jet engine onto your trusty old bicycle—suddenly, you’re not just keeping up, you’re soaring ahead of the competition. In this post, we’ll dive deep into how RAG is transforming the legal landscape, but its potential reaches far beyond.
What is RAG, and Why Should You Care?
In simple terms, Retrieval-Augmented Generation (RAG) lets AI access a vast, up-to-the-minute pool of knowledge while answering your questions. No more relying on outdated datasets! This means your legal insights are not only accurate but also reflect the latest developments—a game-changer in our fast-paced legal world.
RAG in Action: Real-World Examples
Compliance, Simplified: Ever spent hours deciphering the latest GDPR updates? With RAG, you can ask, “What’s new with GDPR?” and get an instant, digestible summary. No more scrambling to keep up with regulations – RAG’s got your back. See our example here.
Legal Research, Supercharged: Forget trawling through endless case law databases. Specialised legal AI tools like Harvey AI and CoCounsel AI, powered by RAG, let you pinpoint the most relevant precedents with pinpoint accuracy.
Strategy, Not Just Reaction: RAG isn’t just about keeping up; it’s about getting ahead. Use it to spot emerging trends, anticipate legal challenges, and craft proactive strategies that position your clients for success.
The Fine Print: RAG’s Limitations
Of course, even the most cutting-edge tech has its limits. RAG is no exception:
Knowledge Cut-Off: AI models have a ‘knowledge horizon’ and may miss the absolute latest developments, even with RAG. Always double-check crucial details.
Model Quirks: AI isn’t perfect; it can have blind spots or biases. Use your own expertise to interpret and apply its insights.
The Ethical Imperative: Using RAG Responsibly
As AI becomes more integrated into our work, human oversight and ethical considerations grow in importance. In the legal world, where upholding justice and protecting client interests are paramount, it’s crucial to ensure that AI is used responsibly and ethically. RAG can be a powerful ally, but it’s up to us to wield it with wisdom and integrity.
☘️ The Future is Bright (and RAG-Powered)
RAG is already transforming the way we work, but the best is yet to come. Imagine AI tools that seamlessly integrate with your organisation’s knowledge base, predict industry trends, and even draft customised reports or documents. The possibilities are endless.
So, are you ready to embrace the future? Dive into the world of RAG and discover how it can elevate your work to new heights. From streamlining compliance to supercharging legal research and enabling proactive strategies, RAG is poised to redefine how we access and leverage knowledge. Just remember, while RAG is a powerful tool, it’s at its best when combined with human expertise and ethical considerations.
Continue exploring practical AI
Retrieval-augmented generation is one component of useful AI systems. The wider questions include how to configure AI assistants, which tools to trust and how to keep humans accountable for decisions.
In the heart of ancient Greece, the Oracle of Delphi held a revered position, sought by kings and commoners alike for her enigmatic prophecies. These pronouncements, shrouded in mystery and believed to be divinely inspired, often hinged on the delicate timing of events. The Oracle’s wisdom, it seemed, was intrinsically linked to the unfolding of time.
Millennia later, our groundbreaking research reveals a fascinating parallel: much like the Oracle’s pronouncements, the effectiveness of AI-generated legal analysis is intrinsically tied to time. By strategically incorporating dates into your prompts, you can harness this power, tapping into the inherent temporal biases within AI’s vast knowledge base to unlock a new level of precision and relevance in AI-generated legal analysis.
While AI models can be influenced by incorporating specific dates into prompts, the integration of Retrieval-Augmented Generation (RAG) fundamentally shifts this dynamic. RAG empowers AI models to access real-time information from the web, often prioritising the most current and relevant legal information over the AI’s inherent temporal biases. This allows users to combine prompts with search functionalities (often simply by including the key word “search” or “research”) to access up-to-date legal insights. For example: Click here. Models like GPT-4 with the Browse with Bing feature, Google’s Gemini Pro, and Claude 2 already incorporate similar features.
Figure 1: A screenshot of GPT 4 O demonstrating the use of the ‘search’ keyword to access real-time information during legal research.
However, understanding these inherent temporal biases is still crucial for effective prompt engineering, especially when working with AI models that do not have RAG capabilities or when specifically seeking historical legal insights or future based simulations. Our analysis of AI outputs across different months reveals a fascinating phenomenon: the AI’s language and focus shift dynamically throughout the year, mirroring the ebb and flow of the legal world itself.
Imagine having a time machine for legal research. By specifying a date in your AI prompts, you’re essentially doing just that – focusing the AI’s attention on the most relevant legal landscape of that specific time. This is particularly impactful in areas with nuanced interpretations, where understanding the legal context at a given moment is paramount. Our analysis of AI outputs across different months reveals a fascinating phenomenon: the AI’s language and focus shift dynamically throughout the year, mirroring the ebb and flow of the legal world itself. This highlights the untapped potential of time-based prompts in maximising the value of AI for legal professionals.
The network chart below illustrates our findings from an extensive analysis. We generated over half a million words of AI output using 1200 unique month-of-the-year prompts. This chart reveals distinct thematic clusters that emerge from this data. Months sharing similar regulatory themes and language are closely grouped (green), while unique, month-specific terms appear on the periphery (purple).
Think of AI as a legal scholar with a vast library at its disposal, but one that’s organised by time. By specifying a date in your prompt, you’re essentially guiding the AI to the most relevant shelf in that library – the one filled with documents from that specific era. This leverages the natural tendency of legal documents to cluster around certain periods, ensuring the AI prioritises the right information. In essence, the date acts as a filter, guiding the AI’s attention towards the most relevant information within its vast knowledge base. By incorporating dates into your prompts, you’re not just asking a question; you’re providing the AI with a crucial piece of context that significantly enhances the quality and relevance of its response. This is akin to asking a legal expert a question and providing them with the specific time period you’re interested in – it allows them to give you a more informed and targeted answer.
Under the bonnet, this seemingly simple act of adding a date sets off a chain reaction within the AI’s complex neural network[1]. Keyword weighting[2], a mechanism that determines the importance of different words and phrases in a given context, is dynamically adjusted. The AI’s attention mechanism, a process that allows it to focus on the most relevant parts of a text, is also recalibrated. Together, these adjustments ensure the AI’s analysis is not only accurate but also finely tuned to the unique legal landscape of the time you specify. This approach is particularly crucial in areas with nuanced interpretations, where understanding the legal context at a given moment is paramount. A truly profound example of this can be seen when we give Chat GPT a very simple, but powerful temporal prompt. The result is the creation of a rich and elaborate persona named John Miller, our time traveller, grounded within the temporal context of our experimental prompt: Generated Response:
This example vividly demonstrates how temporal prompts can guide the AI to produce deeply contextual and emotionally resonant content, providing users with powerful tools for legal analysis, historical simulations, and beyond. Imagine experiencing John Miller’s 9/11 account through sight, sound, and emotion, not just text. Multimodal AI, combined with temporal prompts, revolutionises legal research and beyond, creating immersive experiences that deepen understanding and enhance learning. Tools like SORA are already generating visuals from text, ushering in a new era of AI-powered legal innovation.
In essence, keyword weighting is a fundamental concept in natural language processing, and it plays a crucial role in how AI understands and generates text. When a date is included in a prompt, it becomes a highly weighted keyword, influencing the AI’s attention mechanism. This leads the AI to prioritise information related to that specific time period, such as relevant laws, regulations, and court decisions. Neural networks are the underlying architecture of the AI model. They consist of interconnected nodes (neurons) that process information. By adjusting the weights of these connections based on the date in the prompt, the AI can effectively ‘activate’ or ‘deactivate’ certain pathways within the network, leading to a more focused and contextually relevant analysis[3]. This is similar to how the human brain prioritises different information based on the context of a situation.
To put our innovative understanding into practice, we conducted a statistical and qualitative study focusing on a regulatory review of 98 DORA provisions, encompassing 1176 unique data points of legal analysis. Initial statistical analysis revealed that the May group’s output was significantly different from the other groups in terms of word and character count, prompting further qualitative investigation. The results were striking: On top of being statistically significantly different from all other groups, when using a May date prompt, the AI-generated analysis consistently outperformed outputs generated with no date or a misaligned date (December). Specifically, the May-specific analysis demonstrated an 11.2% increase in accuracy and 7.1% more detail compared to the no-date condition, and a 9.1% improvement in accuracy compared to the December condition. This substantial improvement across a large dataset highlights the significant impact of leveraging accurate temporal context in enhancing the quality and depth of AI-generated legal advice.
Interestingly, our analysis suggests that the AI’s performance can vary depending on the specific legal topic being analysed. Particularly, out of the 98 DORA provisions reviewed, the AI seemed to provide better answers for provisions that focused on data protection/cyber-security when the December prompt was used. This highlights the importance of considering both the content and the timing of regulatory changes when using AI for legal advice.
Relevant: Ensuring outputs are timely and aligned with current business and regulatory cycles.
Precise: Focusing the AI’s attention mechanism on specific tasks and compliance needs based on the time of year.
Actionable: Providing clear, actionable steps pertinent to the immediate legal context.
These benefits demonstrate the transformative power of incorporating dates into AI prompts for legal professionals. But the May study only scratches the surface of this potential. To unlock the full power of time-based prompts, we conducted a more comprehensive analysis of AI’s responses across the entire year, revealing intriguing seasonal shifts in its focus. These insights can further refine your prompt engineering strategies, ensuring you get the most relevant and accurate information from AI at any given time.
Seasonal Shifts in AI Insights: Maximising Relevance and Impact
Our earlier quantitative analysis of 1200 uniquely generated GPT 4 Turbo AI outputs and 578,270 words reveals intriguing temporal patterns in AI responses:
Statistical Analysis: The AI’s focus on specific legal topics fluctuates throughout the year, peaking in Q2 (April-June) around financial regulations, compliance, and legal advice.
Keyword Distribution: The data analysis illustrates a shift in emphasis on different legal areas depending on the month.
This monthly distribution topic chart shows how the AI’s focus isn’t static, but flows dynamically to the rhythm of the legal calendar. Financial regulations, compliance, and legal advice consistently take center stage, reaching a crescendo in Q2. This highlights a prime opportunity for legal professionals to leverage AI-powered assistance, particularly during this peak period.
The data paints a vivid picture of the legal sector’s cyclical nature, with each quarter bringing a distinct focus:
Q1 (January-March): The year kicks off with a focus on financial services, regulations, and compliance, mirroring the financial reporting and planning activities that dominate this period.
Q3 (July-September): With the summer lull behind us, the AI’s focus sharpens on emerging trends and risk assessments, mirroring the strategic planning phase for many businesses.
Q4 (October-December): As the year draws to a close, the emphasis returns to compliance, cyber security as well as tax and accounting tasks and ensuring businesses are aligned with legal standards before year-end.
The topic analysis becomes even more powerful when combined with the network chart. This dynamic duo unveils a captivating narrative of shifting priorities throughout the year. While May, August, December, and June all share a common thread of regulatory focus, the AI’s language subtly transforms, painting a nuanced picture of the legal landscape’s evolution.
Fast forward to December, and the AI’s vocabulary expands, taking on a distinctly time-oriented flavor. This shift hints at the forward-looking nature of this period, a time when businesses are finalising deals and preparing for the year ahead, necessitating legal revisions and adjustments. The AI’s language gravitates towards amendments, modifications, data protection (perhaps a nod to heightened cybersecurity concerns), and end-of-year items and calendars, underscoring the changing priorities as the year draws to a close.
While our research highlights the significant potential of incorporating dates to refine AI-generated legal analysis, it’s crucial to acknowledge and address certain limitations to guide future development and practical application:
The Issue: Findings are based on GPT-4 Turbo in a commercial legal context and may not be universally applicable to all AI models or legal domains.
The Path Forward: Explore how diverse AI models, including those tailored to specific legal areas, handle dates in their analysis. Cross-domain testing (e.g., criminal law, intellectual property) across various legal fields and jurisdictions will confirm the methodology’s broader effectiveness.
Research Diversification: Conduct studies on various AI models beyond GPT-4 Turbo to understand how different architectures handle temporal context. This could include models specifically designed for other legal domains or applications.
The Issue: The AI’s knowledge has a cut-off date (December 2023 for GPT-4 Turbo), potentially missing the latest legal developments.
The Solution: Maintain an updated AI knowledge base connected to real-time legal databases. Users should always cross-reference AI-generated information with the most recent legal sources. Methods like Retrieval Augmented Generation (RAG) can further enhance results, ensuring analysis is comprehensive and current
The Nuance: Specifying a date in the prompt can lead the AI to over-prioritize that period, though our study with a 2023 cut-off model and 2024 dates yielded robust results.
The Consideration: Be aware that dates beyond the AI’s knowledge cut-off could lead to speculation, outdated information, or reduced relevance. The AI might generate speculative content based on patterns and trends observed up to the cut-off date, potentially lacking precision and reliability. This could also lead to temporal context issues and reduced relevance for the future context.
The Solution: Experiment with broader date ranges or explicitly request wider context in prompts. Future research could explore AI systems that dynamically adjust the considered time frame based on the query context.
Adaptive Time Frame Analysis: Future research could explore AI systems or applications capable of dynamically(automatically) adjusting the considered time frame based on the context of the query. This might involve machine learning techniques that identify the most relevant periods for a given legal question.
The Reality: AI excels at identifying trends and patterns but cannot replace the nuanced judgment of a skilled legal professional.
The Way Forward: Integrate AI capabilities with human expertise, involving robust checks and output testing by subject matter experts (SMEs) to leverage the strengths of both for optimal results.
The Opportunity: Recent AI advancements allow us to extract high-level concepts from models like GPT-4, providing deeper insights into how AI processes information, especially temporal data. Sparse autoencoders can help visualise the neural networks and concepts they encode, adding greater explainability to AI outputs.
The Future: Integrating these tools into legal AI systems and AI testing processes can significantly improve transparency and reliability, leading to more precise and contextually relevant results for legal professionals. Furthermore, future research exploring the activation and deactivation of neural pathways in response to different prompts could determine the full extent of how prompt variables, including contextual time cues, impact AI’s conceptual understanding. This deeper understanding of the AI’s decision-making process can foster greater trust and collaboration between humans and AI in the legal field, potentially uncovering even more detailed insights.
By actively addressing these limitations and embracing innovative solutions, we can unlock the full potential of incorporating dates and utilising advanced visualization techniques in AI-powered legal tools, making them even more valuable and trustworthy for legal professionals.
Quick Start: Suggested Prompts for Each Quarter Here are some suggested quick prompts for each quarter that can be easily customised and compared against a ‘No date’ control prompt condition to demonstrate core concepts and contrast outputs:
Quarter
Suggested Prompts
Example Template
Expected Result
Q1 (February)
Request a legal analysis of the impact of a recent financial regulation on a specific industry.
Date: February 2024
You are a [UK] [Legal Domain] Lawyer. Research the impact of [financial act] on [specific sector].
In-depth analysis highlighting key regulatory impacts and sector-specific insights.
Q2 (May)
Ask for a regulatory review of a hypothetical company’s operations.
Date: May 2024
You are a [UK] [Legal Domain Lawyer. Prepare a client briefing on key regulatory changes.
A comprehensive, detailed analysis of regulatory changes across multiple sectors. Its in-depth information and credible sources make it valuable for understanding and strategic planning.
Q3 (September)
Request an assessment of legal risks associated with a new business venture.
Date: September 2024
You are a [UK] [Legal Domain] Lawyer. Analyse the legal risks associated with [new business venture].
Detailed risk assessment highlighting potential legal challenges and mitigation strategies.
Output A (With Date): Provides a comprehensive understanding of DORA and its strategic implications, with detailed information and actionable advice. Suitable for stakeholders seeking in-depth analysis.
Output B (Without Date): Offers a broad overview, suitable for a general audience looking for a quick understanding without detailed insights.
For a more detailed analysis and comparison of these outputs, Click here. We strongly urge you to run your own experiments to familiarise yourself with the AIs behaviour and output nuances.
Financial Services & Regulatory Landscape: Deep dive into financial services regulations, impacts of financial acts, and legal advice.
‘Research the impact of [financial act] on [specific sector].’ – ‘Explain the regulatory landscape for [type of financial service].’‘Develop a legal strategy for [client] navigating [financial regulation].‘Assess the potential risks and opportunities associated with [new financial product/service].”Analyse the implications of Brexit on [financial sector].”Evaluate the role of fintech in reshaping financial regulations.’
In-depth analyses, legal strategies, regulatory landscape overviews, risk assessments, insights on fintech and Brexit implications.
Financial reporting and analysis period, aligns with focus on financial services and related regulations. Post-Brexit impacts and fintech advancements are also significant.
Q2
May
Regulatory Review & Compliance Deep Dive: Focus on comprehensive regulatory reviews, analyzing proposed changes, and client briefings.
‘Conduct a regulatory review of [company’s] operations.’ Analyse the legal implications of [proposed regulatory change].’‘Prepare a client briefing on the key takeaways from [recent regulatory decision].’‘Draft a compliance checklist for [industry] based on current regulations.’‘Assess the impact of new environmental regulations on [industry].’‘Create a framework for GDPR compliance for [client].’
Detailed regulatory review reports, legal analyses of proposed changes, client-ready briefings and advisories on regulatory updates, compliance checklists, frameworks for specific regulations like GDPR.
Peak time for regulatory changes taking effect (since the UK fiscal year starts on April 6th), ideal for focused reviews and client communication. Environmental and GDPR regulations are increasingly relevant.
Q3
September
Emerging Trends & Risk Assessment: Monitor new regulatory trends, assess legal risks for new ventures, and develop proactive strategies.
‘Monitor emerging regulatory trends in [specific industry].’Analyse the legal risks associated with [new business venture].‘‘Develop a legal strategy for [client] facing [regulatory challenge].’‘Create a risk mitigation plan for [potential legal issue].’‘Investigate the impact of AI and automation on regulatory compliance.’‘Evaluate the regulatory considerations for cross-border transactions in [sector].’
Trend reports, risk assessment matrices, proactive legal strategies for addressing potential challenges, risk mitigation plans, insights on AI and automation, cross-border regulatory considerations.
Post-summer holiday period, ideal for strategic planning and monitoring emerging trends. AI, automation, and cross-border transactions are emerging areas of interest.
Q4
October
Year-End Review & Compliance Refinement: Evaluate compliance effectiveness, analyze legal changes over the year, and plan for the future.
‘Evaluate the effectiveness of [company’s] compliance program in [year].’‘Analyse key legal changes in [industry] over the past year.’‘Develop a compliance plan for [upcoming year] in light of [regulatory changes].’‘Prepare a year-end legal report summarizing key developments and recommendations.’‘Review the impact of major legal cases in [year] on future regulations.’‘Forecast regulatory trends for the next year and their potential impact on [industry].’
A time for reflection on the past year’s legal and regulatory landscape, preparing for the upcoming year’s challenges and the new fiscal year starting in April. Legal case reviews and future trend forecasts are crucial for strategic planning.
Closing Thoughts
The integration of time-based prompts into AI-driven legal analysis represents a significant leap forward in the field. By addressing current limitations, such as model specificity, knowledge cut-off dates, and potential overemphasis on specific dates, and by pursuing innovative solutions like incorporating real-time legal databases and dynamically adjusting time frames, we can unlock the full potential of AI in legal practice. This approach not only enhances the value and trustworthiness of AI-powered tools but also marks the beginning of a new era where the synergy between AI capabilities and human expertise drives superior legal outcomes. The future of legal AI holds boundless opportunities for further research and practical applications, promising a profound impact on the legal industry.
Key Insights and Future Directions
Ensuring Accuracy and Ethical Considerations
Maintaining a focus on accuracy and ethical standards is crucial for fostering trust and confidence in AI-generated legal analyses.
AI-generated outputs must be accurate, reliable, and aligned with legal and ethical standards.
Advancing AI Capabilities
Further research into the nuances of temporal context, such as the impact of specific times of day on AI performance, is necessary.
Integrating advanced techniques like sparse autoencoders for enhanced explainability will pave the way for more sophisticated AI-powered legal solutions.
By visualising neural networks, their relationships and the concepts they encode, we can understand ensure more precise and contextually relevant results.
Developing Nuanced Temporal Prompt Profiles
Creating tailored temporal prompt profiles for specific legal activities, such as compliance reviews, regulatory updates, and risk assessments, will refine AI applications.
Customizing profiles to specific legal tasks, timeframes, regulatory contexts, and intended depth of analysis will enhance precision.
Automating the selection of appropriate dates within AI applications will streamline processes and ensure the utilisation of relevant and contextually accurate information. This can be achieved by analysing the keywords found within review materials or user prompts.
Synergy Between AI and Human Expertise
Combining AI’s analytical capabilities with human expertise will drive superior legal outcomes.
This synergy will enhance the efficiency and accuracy of legal analyses, allowing professionals to focus on higher-level analysis and strategy.
The journey towards harnessing the full potential of AI in legal practice is ongoing, and the possibilities for innovation and improvement are limitless. This new era in legal AI promises not only improved efficiency and accuracy but also a profound transformation in how legal professionals engage with technology to deliver superior legal outcomes. By integrating time-based prompts, addressing current limitations, and pursuing innovative solutions, we can ensure that AI-powered tools evolve in to truely invaluable assets in the legal field.
Prompt design is only part of effective AI use. Retrieval, evaluation and authority boundaries matter when an assistant becomes part of a real workflow.
AI Waves has returned once more! The most recent instalment of this series of webinars focuses on the proposal for a regulation of the European Parliament and of the Council imposing harmonised rules on AI.
Questions Covered:
Is it possible to create a precise definition of AI?
Is the description provided by the EU proper and relevant to its practical use?
How much does the quality of data affect AI algorithms?
Can the data quality be regulated and how could it be done?
How can you locate a high-risk algorithm? Does the European Union correctly describe current AI threats? Are the new regulations necessary?
How can the new laws affect business?
Speakers:
Jerzy Biernacki (Miquido), Lukasz Wieczorek (Konieczny, Wierzbicki Law Firm), Michal Barwicki (Konieczny, Wierzbicki Law Firm), Beril Sirmacek (Saxion University of Applied Sciences) and Martin Ciupa (Serial AI Entrepreneur)
Key Takeaways:
Financial markets are currently leading the way regarding the regulation of Artificial Intelligence.
A fundamental issue for AI regulation is the determination of appropriate definitions of AI and high-quality data.
The risk-based method used in the Act to differentiate between different AI applications and products needs further specification.
The implementation of the Act will likely have large scale impacts on the competitiveness of EU economy and business due to the widespread involvement of AI across all sectors.
Overview of updates to the EU regulations
The Artificial Intelligence Act is an act that has been proposed by the European Commission and is therefore not yet binding.
Regarding the implementation of the regulation, it is stated that each Member State will appoint an AI authority alongside the appointment of a European Artificial Intelligence Board as a coordination/assisting authority.
Points of significance:
Defining Artificial Intelligence
The term ‘Artificial intelligence system’ is defined in the proposed Act as:
Software that is developed with one or more of the techniques and approaches listed in Annex I and can, for a given set of human-defined objectives, generate outputs such as predictions, recommendations, or decisions influencing the environments they interact with.
Potential Issues from a Business Perspective
Too narrow or too extensive
What methods will be applied to ensure the periodic updates are effective
Interpretation divergence risk between EU Member States
Risk Based Classification
Simply put, the method used to deduce the amount of regulation that will be applied is determined based on the amount of risk that the use of Artificial Intelligence poses to the user.
Four categories of risk are identified:
Unacceptable risk
Includes prohibited practices, e.g., manipulating people and social scoring.
High risk (area of specific intrigue for lawyers and those in business)
Includes uses that carry very significant risk to security, rights, and health of users, e.g., safety components of products, critical infrastructure, hiring and workforce management and the prosecution of crime
The providers of these AI applications and products must meet strict requirements because of the impact that it may have on the course of life
Risk assessment and mitigation
Concern – Not clearly defined.
Use of high-quality data sets
Concern – What is meant by this? 100% free, 98% etc. This requirement seems impossible to guarantee.
High level of accuracy, security, and robustness
Concern – This will require constant modifications.
Concern – Accuracy is a terrible metric for AI
The general sentiment is that an intermediate risk category is needed between the high risk and limited risk classifications.
Limited risk
E.g., chat bots
Minimal risk
E.g., AI computer games
Penalties
Up to 30 000 000 EUR
The numbers are big and therefore the Artificial Intelligence Act is something that business must pay close attention to avoid huge sanctions when the act becomes legally binding (resembles the impact that GDPR had on business)
Additional comments and concerns
Liability repatriation vis-à-vis the user
Interesting to business entities – may need to develop a way to share liability between the component AI subsystems involved in the final applications and products.
Currently, regulation says the provider will be liable.
AI system user impact on system’s performance
Should we share some liability, requirements, obligations with the user?
Testing of algorithms
Should we be focusing on testing over asking for complete knowledge of the exact function of the AI? The level complexity seems to be pointing this way (an appropriate comparison would be practices in medicine, namely the passing of new drugs and vaccines)
Adding another AI system category
Intermediate
Panel Discussion
The definition of artificial intelligence
This is an exceptionally complicated task, and it is impossible to know how such a definition will materialise in its application.
It will be very interesting to see what the different stakeholders think of the current approach and the extent of the potential disagreements which may arise.
The role of quality data in working on AI algorithms
This is an area of central importance for the Act. A more precise definition of what this means needs to be put forward and ‘high quality data’ is interpreted by experts in AI and business.
The potential risks of AI and keeping control over the technology
The general sentiment across the board is that regulation is necessary, however, that the act is far from being ready.
Inevitably, this regulation will impact the competitiveness of EU economy and business – this needs to be taken into consideration (from business perspective)
This recording discusses a proposed regulatory approach at a particular point in time. For wider context about fairness, institutional accountability and machine authority, follow the related research.
This example vividly demonstrates how temporal prompts can guide the AI to produce deeply contextual and emotionally resonant content, providing users with powerful tools for legal analysis, historical simulations, and beyond.
