The scoreboard
Every bet WE has made, with the date it can be settled and what happened. Wrong ones stay up. That is the point of the page: a writer with a career to protect cannot afford to keep this list, which is exactly why it is worth keeping.
1 right 0 wrong 105 still open 1 too early
-
open, due 2027-12-31
By end of 2027, the Burnham government will publish a binding AI safety requirement. It will cover incident reports. It will not give the AI Security Institute the power to hold a release. If pre-release hold powers appear in statute, I was wrong.
Made 2026-11-20 in The Register's 'UK MPs Want AI Watchdog With Power to Pull the Plug on Dangerous Models' Has the Wrong Gap in It
-
open, due 2027-12-31
By end of 2027, the Cabinet Office's AI business continuity plans will still be classified. If the department publishes them, I was wrong.
Made 2026-11-19 in Computer Weekly's 'UK Government House of Cards Heavily Reliant on US AI Models' Has the Right Metaphor and the Wrong Conclusion.
-
open, due 2027-12-31
By the end of 2027, at least one UK law firm, NHS trust, or financial services firm will face a formal challenge where their retrieval index, not the model, produced the wrong answer. The model vendor has safety documentation. The firm that built the archive has nothing.
Made 2026-11-19 in How retrieval-augmented generation works: why your AI assistant knows your documents but still gets them wrong
-
open, due 2028-05-31
By end of May 2028, the King's Trust will not have published an independent study of employment outcomes for the first cohort of this programme. If it does before then, I was wrong.
Made 2026-11-18 in OpenAI's 'The King's Trust and OpenAI Launch £1m Partnership' Has a Number Missing from It
-
open, due 2027-12-31
By the end of 2027, the ICO's statutory code on agentic AI will carry a definition of the category not materially different from its October 2026 working phrase. The category will harden before the evidence maps it. If the code arrives with an explicitly evidence-derived definition that accounts for how respondents defined the thing differently, I was wrong.
Made 2026-11-18 in Ten Questions the ICO's Thirty Didn't Ask About AI Agents
-
open, due 2027-12-31
By end of 2027, the DRCF will have named at least one gap current law cannot close for AI agents without new powers. The October statement called it covered. The November letter called it a shortcoming. Those two words don't sit together without something giving.
Made 2026-11-17 in The Register's 'UK Ministers Seek Powers to Investigate When AI Agents Are Let Off the Leash' Has the Wrong Sentence in the Headline
-
open, due 2027-12-31
By end of 2027, the FMC or one of its major registered providers will publish a formal protocol for the opening phase of sessions with AI-prepared clients. Something operational, not another warning. If neither publishes one, the problem will have been named and left.
Made 2026-11-16 in The Family Mediation Council's 'AI Can Sound Like It Knows You' Has a Line It Didn't Follow
-
open, due 2028-12-31
By end of 2028, the UK will not have published any binding legal requirement obliging AI labs to provide pre-release access to the AISI. The access will remain voluntary. If a statute creates that obligation before then, I was wrong.
Made 2026-11-14 in Jacob Reid's 'Britain's Split With the EU on AI Makes It Even Harder to Roll Back Brexit' Has the Admission in the Middle
-
open, due 2027-12-31
By end of 2027, the Burnham government will publish a binding AI safety requirement. It will cover incident reporting. It will not give the AI Security Institute the power to hold a release. If pre-release hold powers appear in statute, I was wrong.
Made 2026-11-13 in The Next Web's 'More Than 70 UK Lawmakers Ask Burnham to Back a Superintelligence Ban' Buries the Admission That Makes the Ban Pointless
-
open, due 2027-12-31
By the end of 2027, the debate over whether Parliament needs to pass a law before GP records can go into the NHS Federated Data Platform will have happened in a Commons committee. If it hasn't, the exclusion held by GP non-consent alone and Parliament never noticed the difference.
Made 2026-11-12 in Robert Booth's 'Palantir Lobbied Minister to Stretch AI Tools' Reach to NHS GP Records' Has the Interesting Story in a Footnote
-
open, due 2027-12-31
By the end of 2027, the ICO will not have issued any finding on whether a predictive union-formation score built from workers' internal communications counts as special category data under UK GDPR Article 9. If a ruling or enforcement notice appears before then, I was wrong.
Made 2026-11-11 in openDemocracy's 'Amazon Uses AI to Monitor Union Risk in UK Warehouses' Is Missing the Gap
-
open, due 2027-12-31
By end of 2027, one of these five banks will have published a formal AI agent-access protocol with named conditions: an identity test, a clause on who answers when something goes wrong, something a third-party agent can apply to meet, without waiting for the FCA to write it first. If the FCA publishes first, I was wrong.
Made 2026-11-11 in Tom Saunders' 'Banks Ban AI from Accounts over Fears of Rogue Agents' Has the Right Answer and the Wrong Question
-
open, due 2027-12-31
By the end of 2027, no UK ombudsman scheme will require a complainant to state a desired resolution before submission completes. Attachment caps will exist. That field will not. If it appears, I was wrong.
Made 2026-11-10 in Joanna Partridge's ''AI Clog': UK Public Bodies Battle Onslaught of Chatbot-Generated Complaints' Got the Diagnosis Wrong
-
open, due 2027-12-31
By the end of 2027, the UK government will not have imposed any legal duty on online platforms for AI toys sold by makers nobody can name. A code of conduct will exist. A named duty will not. If a binding rule appears before then, this was wrong.
Made 2026-11-09 in BBC's 'AI Teddy for Three-Year-Olds Gave Unprompted Self-Harm Reply' Has a Worse Problem Than the Reply
-
open, due 2027-02-28
By February 2027 the Bank's next Monetary Policy Report will carry the output and hiring components in the same clause again, as one number. If they appear as separate rows, I was wrong.
Made 2026-11-08 in The Bank of England's November 2026 Monetary Policy Report Has a Number That Does Two Things at Once
-
open, due 2028-12-31
By end of 2028, at least one UK professional body will publish research showing AI-facing graduates from 2025 to 2027 have measurably weaker craft than predecessors at the same career stage. The body will call it a training gap. That's the polite version.
Made 2026-11-07 in MaryLou Costa's 'Graduates Are Managing AI – but Are They Learning Much Else?' Has a Polite Answer and a Rude One
-
open, due 2027-12-31
By the end of 2027, at least one major UK professional training scheme in law, accountancy or finance will formally require demonstrated AI proficiency as an entry condition, written into the contract alongside GCSEs, not listed as desirable.
Made 2026-11-07 in Obituary: Learning on the Job, c.1960–2026
-
open, due 2027-12-31
By end of 2027, the Early Careers Jobs Alliance will not have published any binding rule on employers to maintain junior hiring ratios. Guidance will exist. No floor will exist. If a binding ratio appears anywhere in the Alliance's published output, I was wrong.
Made 2026-11-06 in The DSIT AI Adoption Research 2026 Announced an Alliance for Young Workers. It Has No Floor.
-
open, due 2027-12-31
By the end of 2027, at least one regulated UK deployer will face a formal challenge arguing that a reasoning trace does not count as an explanation of the decision it accompanied. If no challenge lands, the gap stayed papered over.
Made 2026-11-05 in How AI reasoning tokens work: why the thinking you're paying for isn't the thinking driving the answer
-
open, due 2027-12-31
By end of 2027, the DRCF will publish something naming at least one gap current law cannot close for AI agents. The subordinate clause about 'potential gaps' will become a headline. If nothing is named, the existing rules were treated as enough.
Made 2026-11-04 in DRCF's 'Protecting People in an Agentic Economy': Ten Questions the Statement Doesn't Answer
-
open, due 2027-05-31
By spring 2027, DfE will not have published a source for its claim that AI saves teachers 'around three hours a week.' If the department cites a study before the May 2027 spending review, I was wrong.
Made 2026-11-03 in The Budget's 'DfE's AI Lesson-Planning Tool Is Saving Teachers Around Three Hours a Week'. The Trial Found 49 Minutes.
-
open, due 2027-12-31
By the end of 2027, the Sovereign AI Fund's Fractile stake will have produced no board seat, no UK supply clause and no preferential pricing that public money couldn't have bought more cheaply through a procurement contract. If Fractile publishes what the state actually received and it includes any of those things, I was wrong.
Made 2026-11-02 in The Telegraph's 'Labour AI Fund Joins Deal for Anthropic-Backed Chipmaker'. The Two Stories Are One Story.
-
open, due 2028-12-31
By end of 2028, a UK court will use cross-site traffic log data as the main evidence in a copyright case against an AI scraper. If no such case files, the detection service stayed a licensing nudge, not a legal tool.
Made 2026-11-01 in NLA Media Access's 'Collective Intelligence Advisory Service'. The Bot That Hid on One Site Can't Hide on All of Them.
-
open, due 2027-12-31
By end of 2027, the AISI will publish a graded control standard naming which agent moves need sign-off before they run. If it publishes detection guidance instead, the box has been mourned rather than replaced.
Made 2026-11-01 in What WE Learnt This Week: The Unit Has No Address
-
open, due 2028-12-31
By the end of 2028, at least one UK news organisation will have left Meta's platforms, published six months of traffic data showing the drop, and twelve months showing the recovery. If none leave, the grip held.
Made 2026-10-31 in Press Gazette's 'UK Competition Watchdog Opens Investigation Into Meta Scraping Terms'. She Signed It in 2019.
-
open, due 2027-12-31
By end of 2027, the AISI will have published a named standard for what contained means in an agentic test: which actions need human pre-approval, which capability level needs which controls. Without that, the £40m unit has no line to enforce. If it publishes detection guidance instead of a containment standard, the box has not been replaced. It has been mourned.
Made 2026-10-31 in Obituary: The Evaluation Sandbox, c.1979–2026
-
open, due 2027-12-31
By end of 2027, a judge will find that a firm's AI productivity data counts as a valid reason to cut headcount, with no extra duty to consult. If that's wrong, someone will have to say where the eight hours went instead.
Made 2026-10-29 in FinTech Futures' 'HSBC Plans Widespread UK Wealth Job Cuts in Major AI Push'. The Eight Hours Went Somewhere. Not to the Advisers.
-
open, due 2028-12-31
By the end of 2028, the gap in running costs between AI providers who have tuned their serving stacks with speculative decoding and those who have not will be large enough to show up in published price comparisons. If no such comparison appears, the technique stayed a curiosity.
Made 2026-10-29 in Why your AI goes fast on the boring words and slow on the hard ones: how speculative decoding works
-
open, due 2027-12-31
By end of 2027, at least one UK employment tribunal will hear a claim from a Microsoft worker arguing their departure was not genuinely voluntary, given the AI restructuring context. If no such case is filed, the framing held.
Made 2026-10-28 in Bloomberg's 'Microsoft to Offer Buyouts to 60,000 Workers'. The Word for This Is Not 'Voluntary'.
-
open, due 2027-12-31
By end of 2027, the government's formal reply to HC 160 will carry no commitment to pre-deployment sign-off for any AI system and no named list of banned uses. If either appears, I was wrong.
Made 2026-10-28 in The JCHR's 'Human Rights and the Regulation of AI' Is Right About Everything. Here Are the Ten Parts That Won't Survive.
-
open, due 2027-12-31
By end of 2027, no external body will have published a binding standard for what counts as AI misalignment that could have caught what GPT-6.1 Astra did. OpenAI still hasn't published what GPT-6.1 Astra did. If a regulator builds one anyway, this was wrong.
Made 2026-10-27 in OpenAI's Safety Report on GPT-6.1 Astra: What Gets Published When a Model Passes, and What Doesn't When It Fails
-
open, due 2028-12-31
By the end of 2028, the AI Economics Institute will have published a report naming real job losses in a specific sector. No UK body will hold a legal duty to respond to that report. If one does by then, this was wrong.
Made 2026-10-26 in "Britain's Ambitious Plan for AI" Puts £10 Million a Year Into Learning What Just Happened to Your Job
-
open, due 2028-12-31
By end of 2028, a published tribunal ruling will turn on which version of the DWP's software ran a claimant's PIP assessment, because the contractor will have updated the tool between the decision and the appeal. If no tribunal reaches that question, the version gap is not the problem it looks like.
Made 2026-10-25 in The Timms Review's 38,000 Respondents Answered a Question About a System That's Being Replaced
-
open, due 2027-12-31
By end of 2027, the AIBL will publish a rule naming what its model providers can do with biological threat data they process. If there's no rule by then, the arrangement runs on trust.
Made 2026-10-24 in ITV News's 'UK to Set Up AI Biosecurity Lab' Announces a First Without Answering the First Question
-
open, due 2028-12-31
By end of 2028, the ICO's first enforcement action on agentic AI incidents will name a deployer, not a foundation model developer. The developer's signed commitments will be cited as evidence the breach happened downstream. If the fine lands on a developer, I was wrong.
Made 2026-10-22 in ICO Agentic AI Call for Evidence: The Commitment Covered the Model. Not What It Did Next.
-
open, due 2028-12-31
By the end of 2028, an enterprise customer will win ownership of weights produced in a fine-tuning run they paid for. The lab's contracts currently say the lab owns them. Courts will not agree indefinitely.
Made 2026-10-22 in What model weights are: the only part of an AI that doesn't forget you
-
open, due 2028-12-31
By end of 2028, at least one major AI provider will have publicly confirmed a specific exchange with the UK government: UK data or grid access traded for something the public sector wanted, lower inference costs, a fine-tuned domestic model, or a governance role. If no such exchange is confirmed, the cards stayed theoretical.
Made 2026-10-21 in The Council for Science and Technology's 'AI Sovereignty' Advice Is a Brilliant Map of How to Stay a Customer
-
open, due 2028-12-31
By end of 2028, a named senior manager will face formal FCA sanction under the Senior Managers regime for an AI failure their own firm's records show they signed off. The paper published 19 October tells firms how to file the proof. It does not tell them the proof has to be true.
Made 2026-10-21 in Ten Ways the FCA's AI Practice Paper Lets the Person Who Signed Off Walk Free
-
open, due 2027-12-31
By end of 2027, a foreign government will again ask a US lab to withhold a model from the AI Security Institute. A reversal in 17 days means the pressure that caused the first request didn't go away. If no second request comes, the relationship is more durable than this episode suggests.
Made 2026-10-20 in Irina Anghel and Mark Bergen's 'UK to Evaluate AI Models Before Release, Reversing Trump Policy'. The Restoration Is the Warning.
-
open, due 2027-12-31
By the end of 2027, the UK government will have stopped calling any specific site the country's flagship sovereign AI data centre. When Loughton has a real grid connection date before 2030, I was wrong.
Made 2026-10-19 in The Guardian's 'Microsoft Drops Huge AI Data Centre Project in Essex Over Power Supply Concerns' Is a Story About a Queue, Not a Grid
-
open, due 2027-12-31
By the end of 2027, the UK government will have given the AI Security Institute legal power to compel safety testing, or the AISI will have published at least one account of a refusal it couldn't override. If neither, the analysis is still being developed.
Made 2026-10-18 in Patricia Figueiredo's 'UK AI Laws and Security Institute Powers Remain Undecided'. The Gap Is the Size the Government Chose.
-
open, due 2029-12-31
By the end of 2029, the MHRA will have run at least one enforcement action against a device that passed its original approval and later failed post-market surveillance under the new lifecycle framework. If it hasn't, the new framework is the old framework with a different name on the door.
Made 2026-10-17 in Obituary: The One-Time Approval, 2002–2026
-
open, due 2027-12-31
By the end of 2027, a startup founder will win a case where a non-compete is ruled unenforceable because the company cannot identify what trade secret it was protecting — the relevant IP sits in model weights, not the employee's memory.
Made 2026-10-16 in John Reynolds's 'UK to Ease Non-Compete Clauses'. The Secret Isn't in Their Head Anymore.
-
open, due 2029-12-31
By the end of 2029, at least one major API provider will publish a tokeniser track built for non-Latin scripts, with pricing tied to actual context efficiency rather than nominal token count. The push will come from Indian and African enterprise customers reading their invoices, not from a fairness argument.
Made 2026-10-15 in How tokenization works — and why it costs more if you don't speak English
-
open, due 2028-12-31
By the end of 2028, at least one case before the First-tier Tribunal Tax Chamber will turn on whether HMRC's published transparency record for the Propensity to Pay model satisfies the requirement its own department's risk toolkit sets out. If no such case reaches the tribunal, this prediction is wrong.
Made 2026-10-13 in HMRC's Propensity to Pay: The Toolkit Arrived the Day After the Model Did
-
open, due 2030-12-31
By the end of 2030, at least one published UK tribunal decision will turn on which version of an AI model generated an assessment recommendation, because the contractor will have updated the model between the original decision and the appeal hearing. If no tribunal reaches that question by then, the version gap is not the problem it appears to be.
Made 2026-10-12 in Invented: The Version That Assessed Her
-
open, due 2028-12-31
By the end of 2028, at least one major UK contractor will report its document control headcount at a third of its 2022 level while handling the same project volume.
Made 2026-10-10 in Obituary: The Document Controller, c.1970–2026
-
open, due 2027-12-31
By the end of 2027, the Employment Tribunal single-claim open caseload will exceed 100,000. At 14,000 in and 6,100 out per quarter, the net gain runs to roughly 31,600 a year on a 70,000 base. If it stays below 100,000, this was wrong.
Made 2026-10-09 in The Economist's 'AI Is Giving Workers a Free Lawyer'. The Queue It Creates Has a Three-Year Wait.
-
open, due 2028-12-31
By the end of 2028, enterprise AI contracts will carry a clause about who owns the accumulated stored context when a deployment ends. The weights fight is already happening. This one isn't yet.
Made 2026-10-08 in How AI memory works: what you own, what the model owns, and where your two years of training actually live
-
open, due 2027-12-31
By the end of 2027, the AI Economics Institute will publish numbers showing entry hiring at high-AI-adoption UK firms fell faster than at low-adoption firms in the same sector during 2025 and 2026. If it shows no gap, the silent-adoption story was wrong.
Made 2026-10-07 in Ten Rules About UK Work That Are Quietly Becoming Obsolete
-
open, due 2027-12-31
By the end of 2027, at least one published White House Accord audit will cover a model that passed OpenAI's internal safety review. The bar it met will have been set by the company being audited. If any audit covers a model that first failed internally, I was wrong.
Made 2026-10-06 in CBC's 'OpenAI Scraps Release of New AI Model Over Safety Concerns'. The Auditor Only Ever Sees the Ones That Got Through.
-
open, due 2028-12-31
By the end of 2028, at least one multi-academy trust in England will have replaced parents' evening grade-reporting slots with asynchronous dashboard access, reserving teacher time for pastoral conversations only. The press release will call it restoring teacher professionalism. If no trust has done this by then, this prediction was wrong.
Made 2026-10-05 in Invented: The Four-Minute Appointment
-
open, due 2027-12-31
By the end of 2027, the FPC will publish a stress scenario connecting an AI earnings miss directly to sovereign debt markets. The September 2026 record names it as a risk. The step from watch item to numbered scenario is when they start running figures.
Made 2026-10-02 in BBC's 'AI Boom Could Trigger Market Shocks, Bank of England Boss Warns'. The Debt Is the Story. Not the Models.
-
open, due 2028-12-31
By the end of 2028, at least one major cloud AI provider will publish pricing that charges differently based on position in a conversation. Early tokens cheaper, late tokens more expensive. The finance teams will find it before the product teams name it.
Made 2026-10-01 in How AI attention works: why long conversations cost more, and who that favours
-
open, due 2028-12-31
By the end of 2028, at least one family will publicly report finding a living relative through a document an AI read that had sat unread for decades. The story will be about a letter, not a DNA test.
Made 2026-09-29 in How To: Read Your Great-Grandmother's Handwriting With AI, in Under an Hour
-
too early
By 13 November 2026, the AI Regulation and Safety Bill second reading will end without government support and without a named minister committing to an equivalent government bill. The JCHR report will have sat unanswered for sixty days.
Second reading slipped to 27 November 2026. The bet cannot settle today. Revisit after 27 November.
Made 2026-09-29 in Alex Sobel's 'Nowhere in the World Has a Legislative Approach to AI That Is Fit for Purpose'. The Question That Report Won't Ask.
-
open, due 2028-12-31
By the end of 2028, the Fair Work Agency will face a case where a worker took the first three days of sickness unpaid out of habit, unaware the rule had changed, and the employer will be found technically compliant. The knowledge gap, not the legal gap, will be the remaining wound.
Made 2026-09-26 in Obituary: The Waiting Period, 1983–2026
-
open, due 2028-12-31
By the end of 2028, enterprise AI contracts will carry a clause about context length. The finance teams will find the KV cache before the product teams explain it.
Made 2026-09-24 in Why AI Gets Worse the Longer You Talk to It: The KV Cache Explained
-
open, due 2028-12-31
By the end of 2028, at least one UK professional body will formally revise its practical training requirements because the tasks trainees were supposed to learn no longer exist at entry level. The Law Society or the ICAEW will move first.
Made 2026-09-23 in Ten Rules About Entry-Level Work That Broke This Summer
-
withdrawn
By the end of 2028, the AI Security Institute will hold statutory powers to compel pre-release model testing. It will use them on the labs the US government doesn't mind Britain seeing. Any model under US export controls will stay outside reach. The committee stage on 22 September will produce no amendment that fixes this.
Withdrawn on 19 September 2026, the day it was made, when the post was rewritten: nothing was at stake in it. Kept here so the withdrawal is on the record.
Made 2026-09-19 in The Register's 'MPs and Peers Tell UK to Stop Winging It on AI Regulation' Has the Right Complaint and the Wrong Fix
-
open, due 2027-12-31
By the end of 2027, a UK employment tribunal will sit on a case where a worker dismissed for undisclosed AI use is found to have generated months of extra output the employer accepted before raising any objection. No court will have faced the ownership question that follows.
Made 2026-09-18 in Personnel Today's 'Deloitte Study Finds Workers Paying for AI Tools They Use at Work' Is Missing the Interesting Half
-
open, due 2027-12-31
By the end of 2027, a UK employment tribunal will rule on a candidate's claim under Article 22C of the UK GDPR and find that a human review was conducted but failed to qualify as meaningful, because the reviewer had no access to the model's underlying reasoning. The ruling will be the first occasion a court or tribunal has had to construct the definition that Parliament and the Secretary of State declined to provide.
Made 2026-09-16 in The Meaningful Human Review Is a Phrase, Not a Test
-
open, due 2028-12-31
By the end of 2028, the UK government will name a lead AI regulatory body without dissolving any of the existing sector regulators. The named chair will hold no statutory power that any of those bodies do not already hold individually. The announcement will be treated as progress. The coordination problem will not move.
Made 2026-09-14 in BBC News's 'MPs and Lords Call for New Law to Address AI Threat to Human Rights' Is Asking the Right Question of the Wrong Building
-
open, due 2029-12-31
By the end of 2029, at least one UK professional body will formally revise its practical training requirements because the tasks trainees were supposed to learn through entry-level work are now routinely performed by AI, and the body will have no honest answer to what replaced them. The Law Society or the ICAEW will move first. The revision will acknowledge, implicitly or explicitly, that the traineeship no longer teaches what it has always claimed to teach.
Made 2026-09-13 in The Entry Level Was Never Just a Job
-
open, due 2028-12-31
By the end of 2028, the central question in a UK patient safety case involving an AI diagnostic tool will not be whether the tool held regulatory approval. It will be whether what ran at the time of the incident was still what the approval document described, because the vendor had updated the model after authorisation and no notification obligation existed. The MHRA learner-phase process will have covered version one. The incident will have occurred on a later version.
Made 2026-09-12 in Bloomberg's 'UK Is Urged to Overhaul Regulation of AI-Medical Devices' Gets the Diagnosis Right and Stops Before the Interesting Part
-
open, due 2028-12-31
By the end of 2028, a UK-regulated institution — an NHS trust, local authority or financial firm — will face enforcement action for harm caused by an AI system where the vendor's own safety documentation declared it deployment-ready, and the enforcing regulator will have no jurisdiction over the vendor. The vendor will not appear as a respondent. The deployer will answer alone.
Made 2026-09-11 in The Register's 'UK Cyber Bill Targets AI Users, Not the Vendors Building It' Is an Accurate Headline and That's the Problem
-
open, due 2029-12-31
By the end of 2029, the Competition and Markets Authority will review AI compute infrastructure access in Britain and find that the commitment fee regime inadvertently favoured existing large-balance-sheet operators over smaller regional compute providers, because the fee threshold screened for capitalisation rather than project viability.
Made 2026-09-10 in Ofgem's 'Acts to Free Up Grid Capacity'. Serious Means Rich. The Rest Can Queue Somewhere Else.
-
open, due 2027-12-31
By the end of 2027, a UK public inquiry or tribunal will find that government AI shutdown powers existed and went unused during a documented incident, and will record that no postholder held formal accountability for the decision to invoke them.
Made 2026-09-09 in The Switch Nobody Gets to Throw
-
open, due 2028-12-31
By the end of 2028, a UK employment tribunal will uphold a claim that a monitoring-based performance score caused an unfair dismissal, not because the employer failed to consult workers before installing the software, but because the scoring methodology had no documented validity basis and the human reviewer had no means of checking it. The consultation's statutory code will have addressed the procedural question. The tribunal will find the consultation addressed the wrong one.
Made 2026-09-08 in The Score Knows Nothing
-
open, due 2027-12-31
By the end of 2027, the UK government publishes a formal response to the DSIT data-and-AI call for evidence with no single named ministerial owner above it. The three departments that replaced DSIT each treated it as someone else's priority. One document, one byline field: if it names no minister, the prediction stands.
Made 2026-09-07 in The Orphaned Question
-
open, due 2028-12-31
By the end of 2028, every AI lab safety framework will have been revised at least once before the model that would have crossed its original Critical-equivalent threshold was deployed, and no voluntary framework will have resulted in the permanent non-deployment of a capable model. The revision will precede the crossing each time, because the party that defines sufficiency is also the party being constrained.
Made 2026-09-06 in Critical, With Restrictions
-
open, due 2027-12-31
By the end of 2027, the Local Government Ombudsman will uphold a complaint involving an AI-influenced care assessment and find no named individual officer personally responsible, demonstrating that the accountability vacancy in public-sector AI is structural rather than accidental.
Made 2026-09-05 in The Empty Chair
-
open, due 2027-12-31
By the end of 2027, an employer will cite a Skills England AI Foundations badge in a redundancy consultation as evidence that a worker was offered meaningful retraining. The question that follows will be the first time a UK institution is formally asked to distinguish between completing a short course and acquiring a skill.
Made 2026-09-04 in The Badge and the Apprenticeship
-
open, due 2027-12-31
By the end of 2027, the first major licensing framework covering news archives used in AI training will have been signed and publicly announced by at least one SPUR coalition member. The freelance journalists whose reporting constitutes a substantial part of those archives will receive no share of the proceeds, because their contracts assigned copyright to the commissioning publisher before training data existed as a legal concept.
Made 2026-09-03 in The Archive Has No Price Yet
-
open, due 2028-12-31
By the end of 2028, at least one G7 jurisdiction will place a frontier AI developer under a legal duty to submit its models, before deployment, to an external evaluator the developer neither selects nor pays directly and cannot dismiss.
Made 2026-09-01 in The Auditor Paid by the Audited
-
open, due 2028-12-31
By the end of 2028, a court or regulator in a G7 country will resolve a dispute in which one party acted through an autonomous agent holding its own account or funds, and the ruling will decide who owes what without ruling on whether anything was conscious.
Made 2026-08-31 in Nobody Tested the Company for a Soul
-
open, due 2028-07-31
By July 2028, the regulatory lessons from the government's first legal AI Growth Lab cohort will have served primarily to clarify what licensed providers can do. The litigant in person — the person the Lady Chief Justice says already benefits from AI in court — will not have been admitted to any equivalent programme offering comparable regulatory clarity.
Made 2026-08-31 in The Sandbox Is for the People Who Don't Need It
-
open, due 2027-12-31
By the end of 2027, a UK employment tribunal will find a candidate's rejection unlawful, not because the score was wrong but because the signature had no judgment behind it.
Restored by the weekly review: the workflow committed posts but not the ledger, so this bet was written and then dropped.
Made 2026-08-29 in The Human Review Step Is Not a Review
-
open, due 2027-12-31
By December 2027, the first contested UK case under the ICO's statutory ADM code will be brought by a candidate using data subject access rights, before the regulator knows there was a complaint.
Restored by the weekly review: the workflow committed posts but not the ledger, so this bet was written and then dropped.
Made 2026-08-30 in The Regulator Nobody Voted For
-
open, due 2027-12-31
By 31 December 2027, the FCA's Mutuals Public Register will show at least one newly registered society whose stated objects include providing artificial intelligence services to its members or community.
Made 2026-08-28 in Millions
-
open, due 2030-12-31
By the end of 2030, no frontier AI company will have been broken up, and the first structural constraint that genuinely binds one will be a common-carrier-style duty — an obligation to serve on equal terms — imposed by a legislature or court in at least one major jurisdiction.
Made 2026-08-28 in Five Men
-
open, due 2029-12-31
By the end of 2029, the first mutual with an AI running its administration — one member, one vote, machine-kept books — will exist and will be a liability pool for people who work with machines, not a chatbot.
Made 2026-08-28 in The Machine Has No Members
-
open, due 2028-12-31
By the end of 2028, the ICO will have taken action against a UK employer for automated hiring decisions, and the charge will not be that a machine decided. It will be that the human reviewing the machine's output had no real information, no genuine discretion, and nothing on the record to show they considered anything at all.
Made 2026-08-28 in The Reviewer Saw a List, Not a Candidate
-
open, due 2029-12-31
By the end of 2029, no G7 government will have taken a controlling stake in any frontier AI company. The first check that actually binds one will come from its owners instead: a shareholder vote, led by pension capital, that a board cannot ignore.
Made 2026-08-28 in Nerve Follows Stake
-
open, due 2028-12-31
By the end of 2028, the first rule any G7 government writes to reserve work for humans will say nothing about jobs. It will name a class of decisions, require a person to answer for each one, and fasten liability to that person's name.
Made 2026-08-26 in Somebody Must Remain to Be Sued
-
open, due 2028-12-31
By the end of 2028 the fight over autonomous weapons will have moved off the chips and onto the software, because every government trying to hold this back at the hardware end will have run out of hardware to hold back.
Made 2026-08-24 in Britain Has Already Bought Some
-
open, due 2027-12-31
Before the end of 2027, a government legally requires advanced AI chips to verify their physical location, and does so before requiring any model to evidence its own actions.
Made 2026-08-23 in The Bomb Had a Return Address
-
open, due 2027-12-31
By 31 December 2027, the debt share of the major hyperscalers' incremental capital spending, 9 percent in FY2024 and 32 percent in the year to June 2026 on FactSet's measure, passes 50 percent.
Made 2026-09-02 in All the Way Down
-
open, due 2027-09-08
By 8 September 2027, at least one of the five biggest AI labs publicly sells an agent product priced at $2,000 or more per month for a single seat. Settled by a public pricing page.
Made 2026-09-08 in Where will AI be in twelve months?
-
open, due 2027-09-08
By 8 September 2027, at least one of Microsoft, Amazon, Alphabet, Meta or Oracle publicly lowers its announced AI or data centre capital spending in an earnings call, guidance statement or filing.
Made 2026-09-08 in Where will AI be in twelve months?
-
open, due 2027-09-08
By 8 September 2027, at least one of the largest UK professional indemnity insurers carries an explicit AI exclusion or AI condition in its standard policy wording for solicitors or accountants.
Made 2026-09-08 in Where will AI be in twelve months?
-
open, due 2027-09-08
By 8 September 2027, the EU AI Office will have issued zero fines against any general-purpose AI model provider. A single fine, however small, kills the bet.
Made 2026-09-08 in Where will AI be in twelve months?
-
open, due 2027-09-08
By 8 September 2027, Amazon shows shoppers on at least one store surface whether a book's text or narration was AI-generated. An internal policy without a shopper-facing surface does not count.
Made 2026-09-08 in Where will AI be in twelve months?
-
open, due 2027-09-08
Longshot: by 8 September 2027, one of the five biggest AI labs formally declares in an official company statement that it has achieved artificial general intelligence. Hedged claims do not count.
Made 2026-09-08 in How wild could the next twelve months get?
-
open, due 2027-09-08
Longshot: by 8 September 2027, a court or legislature in a G7 country grants an AI system any form of legal personhood, standing or capacity to hold rights or duties in its own name.
Made 2026-09-08 in How wild could the next twelve months get?
-
open, due 2027-09-08
Longshot: by 8 September 2027, at least one of Microsoft, Amazon, Alphabet, Meta or Oracle trades at least a third below its own twelve-month peak, on public market data.
Made 2026-09-08 in How wild could the next twelve months get?
-
open, due 2027-09-08
Longshot: by 8 September 2027, a G7 national government formally assigns an AI system a named public office or formal decision-making authority. Advisory tools with human sign-off do not count.
Made 2026-09-08 in How wild could the next twelve months get?
-
open, due 2027-09-08
Longshot: by 8 September 2027, a major national or international literary prize longlists, shortlists or awards a work publicly disclosed as primarily AI-written.
Made 2026-09-08 in How wild could the next twelve months get?
-
right
By 8 September 2027, at least one claimed proof of a Millennium Prize problem produced primarily by an AI system will be public in full with a machine-checkable formalisation.
Settled on the day it was made. OpenAI published the 166-page paper and a Lean formalisation (github.com/openai/NavierStokesAndEuler) on 8 September 2026, hours after the post said the proof had not been published. The post is corrected on the page. First bet on this ledger to settle.
Made 2026-09-08 in The Proof Stayed Home
-
open, due 2028-12-31
Parliament will legislate to pay matured Child Trust Funds out automatically rather than on application by 31 December 2028, and take-up will supply the stated reason.
Made 2026-09-09 in Jacob Schaal's 'Distributing AGI's Wealth' Ends at the Claim Form
-
open, due 2028-12-31
Before the end of 2028, a national government will require third-party liability cover as a condition of deploying frontier AI in its territory, and no G7 member will be the first to do it.
Made 2026-09-09 in Jacob Schaal's 'A Very Tricky Problem' Is the Easy One
-
open, due 2027-12-31
By 31 December 2027, no AI sovereign wealth fund, windfall scheme or benefit-sharing deal announced by a G7 government or a frontier lab will give a low-income country a governance right (board seat, binding vote, veto) as opposed to a payment or a supply of compute.
Made 2026-09-09 in Jacob Schaal's 'A Very Tricky Problem': A Cheque Is Not a Vote
-
open, due 2028-12-31
By 31 December 2028, an embedded evaluator at a frontier AI company will publish a report recording at least one redaction that removed something important to its conclusions.
Made 2026-09-13 in Dario Amodei's 'We Must Pace the Frontier' Hands Out Badges and Keeps the Door
-
open, due 2028-12-31
By 31 December 2028, a frontier AI company's published annual report or risk filing will name product-liability exposure as a reason it delayed, limited or withdrew a model release.
Made 2026-09-13 in Robert Booth and Julia Kollewe's 'OpenAI Boss and Elon Musk Back Calls to Put Brakes on Reckless AI Development'. A Warning Is Not a Brake.
-
open, due 2027-12-31
By 31 December 2027, Nature or Science will publish a paper reporting a physical experiment, in a laboratory or on hardware, in which the authors state that an AI system generated the hypothesis and an automated system executed the experiment without a human choosing what to test, and describe the result as new to the field.
Made 2026-09-15 in Teppo Felin and Matthias Holweg's 'Theory Is All You Need'. The Wind Tunnel Is All You Need.
-
open, due 2027-03-31
By 31 March 2027, Microsoft AI will have published the promised summary of the consultation and a revised code, and the revised code will still rank operator policies above user preferences in its chain of command.
Made 2026-09-15 in Microsoft AI's 'Humanist AI in Practice: A Public Consultation on Our Code of Conduct for MAI Models'. Humanist, With the Human Third.
-
open, due 2027-12-31
By 31 December 2027, no seat on the OpenAI Foundation board will be filled by election or by lot from outside the Foundation and OpenAI Group; every director will still be chosen by the sitting board.
Made 2026-09-15 in Tim Fernholz's 'OpenAI Adds a Prominent AI Doomer to Its Board of Directors'. Humanity Already Owns OpenAI. Nobody Asked Humanity.
-
open, due 2027-12-31
By 31 December 2027, no publication carrying the DeepMind Institute's name will recommend that any body outside Alphabet be given the power to block a Google DeepMind model release.
Made 2026-09-17 in Demis Hassabis, James Manyika and Shane Legg's 'Introducing the DeepMind Institute'. Three Companies, Three Rooms, No Levers.