AI Summaries Rewrite Memory: Who Answers When the Recap Is Wrong? #
A red car rolls towards a junction. A pedestrian steps into the road. The car turns, hits the pedestrian, who goes down, gets up and walks over to the driver. The whole thing lasts 25 seconds, is rendered in the engine of Grand Theft Auto V, and is about as simple as evidence gets. One thing happens in it, and that thing is a collision.
When researchers asked two consumer AI systems to write factual, neutral summaries of that clip and a near-identical twin, 19 of the 20 summaries they got back failed to mention that the car hit the pedestrian.
That finding comes from a paper posted to arXiv on 23 September 2026, “AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory”, by Mattea Sim of Georgetown University, Yael Eiger of the University of Washington and Tadayoshi Kohno of Georgetown. It was picked up over the following days by The Independent, Neuroscience News and a string of science outlets, and it is due to be presented at the AAAI/ACM Conference on AI, Ethics and Society in October. The headline numbers travelled fast. The machines were bad at summarising. And when people read a summary that got a detail wrong, their memory of what they had seen with their own eyes bent to fit the text.
That second finding is the one worth sitting with. Copilot writes up the meeting you sat through. Your phone condenses the news alerts you swiped past. Police officers in several American states file reports drafted by software from body camera audio. Each tool is sold as a memory aid. The study suggests that when the aid is wrong, it does not merely fail to help. It can overwrite.
So what happens to a person's own memory when the machine's recap of their experience is wrong? And when a summary quietly edits the past someone believes they lived, who is supposed to catch it: the reader, the company that sold the recap, or, as things currently stand, more or less nobody?
Twenty Summaries and One Missing Collision #
The researchers used two short animated clips that already had a history in memory research. Both show a red saloon approaching an intersection and striking a pedestrian. The only meaningful difference is the road sign at the junction: in one video it is a stop sign, in the other a yield sign (the American equivalent of a give way sign). The team fed each clip to two models, OpenAI's ChatGPT-5.5 and Google's Gemini 2.5 Flash-Lite, five times per model per video, using a single standardised prompt. The instruction asked for a factual and neutral summary of at least 300 words, with no speculation and no invented facts. That gave them 20 summaries to analyse.
Every one of the 20 contained mistakes, somewhere between seven and 21 per summary. All of them left things out. According to the paper, 65 per cent contained what the authors classed as critical inaccuracies, 90 per cent contained non-critical inaccuracies, and 60 per cent added material that simply was not there. The summaries hallucinated extra pedestrians and extra vehicles and described erratic driving the video did not show. One stated that the pedestrian slightly repositioned but did not enter the roadway, when the pedestrian plainly walks into the road.
Averaged across all prompts and both models, the summaries omitted 51.6 per cent of the central details of the event. And in 95 per cent of cases, the summary did not mention the central detail at all: the car colliding with the pedestrian.
Summaries are supposed to be lossy, but their entire job is to keep what matters and drop what does not. These models did the reverse, producing long, fluent paragraphs about a scene in which the one thing that happened did not happen.
“I was struck by how bad the summaries were, even at this stage in AI development,” Eiger said in the coverage published by Neuroscience News.
Some caution is due. This was a small corpus: two short animated clips, two models, one prompt, 20 outputs. The authors say so plainly in their limitations section, acknowledging that one standardised prompt and a modest number of summaries limit how far the results generalise. Models change month to month, so nobody should read these figures as a stable error rate for all AI video summarisation. What they do show is that two widely available consumer systems, given a clear and simple task, were confidently and substantially wrong about the most important fact in front of them.
Fifty Years of Stop Signs and Yield Signs #
The second half of the study is where the story turns from software quality to human cognition, and where it plugs directly into one of the most studied findings in psychology.
In 1974, Elizabeth Loftus, then at the University of Washington, and her colleague John Palmer showed people films of car accidents and asked them how fast the cars were going. The trick was in the verb. Asked how fast the cars were going when they “smashed” into each other, participants gave higher estimates of speed than those asked about cars that “collided”, “bumped”, “hit” or “contacted”. A week later, people who had heard “smashed” were more likely to say yes when asked whether they had seen broken glass. There was no broken glass in the film. The paper became one of the founding texts of what is now called the misinformation effect.
Four years later, Loftus, David Miller and Helen Burns published a series of experiments in the Journal of Experimental Psychology that will look very familiar to anyone who has just read the Georgetown and Washington paper. Across five experiments and a pilot, 1,242 participants watched a sequence of slides showing an accident between a car and a pedestrian at a junction. Afterwards some were fed information consistent with what they had seen, and some were fed misleading information: a stop sign became a yield sign, or the other way round. Misleading information made people less accurate on later recognition tests, and it had a larger effect when introduced just before the test than immediately after the event. The authors concluded that information a witness encounters after an event, consistent or misleading, gets integrated into the witness's memory of it.
The new study is, in effect, that experiment rerun for the age of the chatbot, split into two sessions between 24 and 48 hours apart. In the first, people watched one of the two animated clips. In the second, they read a summary of what they had seen, then answered a 12-item recognition test in which each question offered two choices. The critical question asked whether the car had approached a stop sign or a yield sign.
Here is a crucial design detail that some coverage skated past. The misleading summaries were not raw machine errors. The researchers took AI-generated summaries and edited them, removing sentences describing the details that differed between the two clips and splicing in AI-generated text from other outputs, so that participants who had seen a stop sign read about a yield sign and vice versa. In other words, the memory experiment tested what happens when an AI-style summary contains a specific, plausible error. It did not test whether the particular errors the models spontaneously made would distort memory in the same way.
That caveat matters, but it cuts less deeply than it seems. The first half of the study showed that errors of exactly this kind are routine. The second half showed what they do once they land. Of the 331 participants who completed both sessions (328 after attention checks), those who read a summary consistent with the video answered the critical sign question correctly 83.6 per cent of the time. Those who read a misleading summary got it right 44.8 per cent of the time. On a two-choice question, that is worse than a coin toss. The statistics were not subtle: the probability of a gap that size arising by chance was below one in a thousand.
Meanwhile participants remembered other details well. On a non-critical question about whether the scene took place in daytime or at night, accuracy was 96.6 per cent. People had not forgotten the video. They had swapped one piece of it for the version they had been told.
“AI is a new method of delivering misinformation, and it has the potential to create these false memories for people who are reading that information,” Sim said in the Georgetown announcement.
Why Knowing It Was a Machine Did Not Help #
Participants were randomly told that the summary they read had been written either by an AI or by a human. In reality all of the summaries were AI-generated. If a machine label worked as a kind of cognitive inoculation, you would expect people who knew they were reading AI output to discount its claims and lean more on their own memory.
They did not. Among those who read a misleading summary labelled as AI-generated, 46.3 per cent answered the critical question correctly. Among those who read the same kind of summary labelled as human-written, 43.4 per cent did. The difference was not statistically significant. The researchers also measured participants' trust in AI and how much they used it, and neither made a meaningful difference to how susceptible they were.
This is consistent with what the misinformation literature has long suggested about warnings. In a 2005 review in Learning and Memory, looking back on three decades of this research, Loftus wrote that warnings about the potential for misinformation sometimes work to inhibit its damaging effects, but only under limited circumstances. The Georgetown and Washington authors reach a similar conclusion. They suggest that noticing discrepancies between the original event and later information can reduce the impact of misinformation, and float warning people about potential discrepancies before they read an AI summary. But they also note that simply knowing the source was AI provided no protection in their data.
Related evidence points the same way. In 2024 a team at the MIT Media Lab, with Loftus among the co-authors, had 200 participants watch a crime video and then answer questions, five of them misleading, via a survey, a scripted chatbot or a generative chatbot built on a large language model. The generative chatbot induced more than three times as many immediate false memories as the control condition, and a week later those false memories persisted, held with higher confidence than in the control group.
A fluent system can introduce errors into your memory. Telling you it is a fluent system does not reliably stop that. And the errors can outlast the conversation that planted them.
The Meeting You Attended, as Told by Copilot #
Move from a laboratory junction to an office, and the stakes change shape without shrinking.
Automated meeting recaps are now a standard feature across the major collaboration platforms. Microsoft's intelligent recap in Teams, available with Teams Premium or a Copilot licence, generates AI summaries and follow-up tasks, marks when your name was mentioned, organises the discussion into chapters, and tells you who said what. Microsoft's own support documentation is candid about the risk: “AI-generated content is based on the event transcript and is powered by GPT and other AI models. The AI-generated content may occasionally be inaccurate, incomplete, or inappropriate.”
That is a reasonable disclosure. But if an AI label did not protect participants from absorbing a planted error about a road sign, there is little reason to think small print will protect an employee from absorbing an error about who agreed to what in a budget meeting.
The people most likely to rely on a recap are those who were least attentive: double-booked, multitasking, half-listening while answering email. They are witnesses with weak original traces, exactly where post-event information has most room to take hold. And if a recap is reread just before someone needs to remember, the 1978 finding suggests that is exactly when misleading information bites hardest. A recap consulted before a Friday dispute is not merely a record of Monday's meeting. By Friday, it may be the meeting.
A change Microsoft has scheduled for this autumn sharpens the point. In a Microsoft 365 message centre notice first published in April 2026, the company announced a “recap without saving transcript” option, which lets Copilot users generate an AI meeting summary from live meeting context without retaining a recording or transcript. The notice says the feature addresses scenarios where organisations want AI-generated recaps “but have policies that limit or prohibit the retention of transcripts or recordings”, and it advises administrators to review their AI governance policies to decide whether it fits their compliance posture. Rollout is scheduled for October 2026.
There are sound privacy reasons for not keeping transcripts. But if the summary is the only artefact that survives the meeting, there is nothing to check it against except the attendees' memories, and the study is precisely about what happens to those memories after people read a summary that is wrong.
Even the transcript layer is not safe ground. In October 2024 an Associated Press investigation reported that OpenAI's Whisper speech recognition model was prone to inventing chunks of text, and that more than 30,000 clinicians and 40 health systems were using a Whisper-based tool from Nabla to transcribe and summarise patient visits, despite OpenAI's warnings against use in high-risk domains. The layer that turns a live event into text can introduce words nobody said.
News You Only Half Read #
For most people, the summary that shapes memory is not a meeting recap but a news digest glimpsed on a lock screen. In December 2024 and January 2025, Apple Intelligence's notification summaries produced a run of false headlines attributed to news organisations. One summary of BBC News alerts suggested that Luigi Mangione, the man charged over the killing of the UnitedHealthcare executive Brian Thompson, had shot himself. He had not. Other summaries wrongly claimed, under the BBC's name, that the darts player Luke Littler had won the PDC World Championship before the final had been played, and that Rafael Nadal had come out as gay. A summary of New York Times alerts said Israel's prime minister, Benjamin Netanyahu, had been arrested. The BBC complained. In January 2025 Apple d notification summaries for news and entertainment apps in a software update, and said summaries would be shown in italics to distinguish them from ordinary notifications. In July 2025 the iOS 26 beta brought news summaries back, this time with a warning beneath the setting: “Summarization may change the meaning of the original headlines. Verify information.” Given the Georgetown and Washington finding that knowing a summary came from AI did not protect memory, a label is a thin defence.
Those were spectacular, checkable errors. The quieter problem is the everyday rate. In February 2025 the BBC published research in which its journalists put 100 news questions to ChatGPT, Microsoft Copilot, Google Gemini and Perplexity, asking them to use BBC sources. It found that 51 per cent of answers had significant issues, 19 per cent of answers citing BBC content introduced factual errors such as wrong dates and figures, and 13 per cent of quotations attributed to BBC articles had been altered or did not exist in the cited article. In October 2025 a far larger follow-up, coordinated by the European Broadcasting Union and led by the BBC, assessed around 3,000 responses across 14 languages and found that 45 per cent of AI answers about news had at least one significant issue, with sourcing the most common problem.
Now put those error rates next to the audience. The Reuters Institute's Digital News Report 2026, published in June, found that weekly use of AI chatbots for news had risen to 10 per cent of people across the markets it surveys, and to 17 per cent among 18 to 24 year olds. Getting summaries was among the most common uses, cited by about a third of those using chatbots for news, and only a small fraction of users regularly clicked through to the original source.
That last number is the hinge. The traditional corrective for a bad summary is the source it summarised, but the whole value proposition of a summary is that you will not open the source. A reader who half-reads an AI digest of a court case and never opens the report carries a version of events that, on this evidence, will not simply sit alongside the truth. For memory's purposes, it becomes what they recall.
Google's AI Overviews were mocked in May 2024 for suggesting glue on pizza and eating rocks, and Liz Reid, Google's head of Search, acknowledged that some answers had drawn on satirical or joke content. Those errors were funny because they were absurd. The errors that reshape memory are not absurd. They are a yield sign where there was a stop sign.
Police Reports and the Officer in the Loop #
The authors' main concern is policing, where the question of who checks the machine is most concrete. Axon, the largest supplier of police body cameras in the United States, sells a product called Draft One that uses generative AI to draft police report narratives from body-worn camera audio. The officer is expected to review and edit the draft before submitting it. In July 2025, the Electronic Frontier Foundation published an investigation arguing that the product was designed in a way that frustrated auditing. According to the EFF, Draft One does not save the draft it generates or subsequent edited versions, so once an officer exports the final report there is no record of which parts were written by the AI and which by the officer.
Put that alongside the new study. The standard defence of tools like this is the human in the loop: the officer was there, knows what happened, and will correct the machine. The paper's authors address exactly that defence. “Though 'humans-in-the-loop' are often expected to correct for AI's mistakes,” they write, “our work suggests human memory can instead be distorted by these mistakes.”
The logic is brutal. The officer is the check on the draft, but reading the draft is precisely the kind of post-event exposure that alters memory. If it says the pedestrian stepped off the kerb against the signal, the reviewing officer may come to remember it that way. The check has been contaminated by the thing it was meant to check, and if the draft is deleted, nobody can reconstruct which version came first.
In the Georgetown announcement, Eiger said it worried her that police departments may be using video summarisation tools without rigorous testing, and without an awareness of how wrong AI-generated summaries could be.
Legislators in two American states have begun to respond. Utah's SB 180, which took effect in May 2025, requires police reports produced wholly or partly with generative AI to carry a disclaimer and requires the author to certify that the report was checked for accuracy. California went further. SB 524, signed by Governor Gavin Newsom in October 2025 and in force since 1 January 2026, requires disclosure of AI use on reports, requires agencies to retain the first draft for as long as the final report is kept, and bars vendors from selling or sharing information police provide to the AI. As the EFF put it, the draft retention requirement means judges, defence lawyers or auditors can see which portions of the final report were written by the officer and which by the computer.
Notice what these rules assume. Utah makes the officer answerable for accuracy, yet the officer's memory is one of the things the draft may have altered. California's retention rule is more robust because it does not rely on memory at all: it preserves an artefact that can be compared against the footage. That is a useful template elsewhere.
Records That Outlive the Room #
Outside policing, the law has not caught up with the idea that a summary can shape the memories of the people it describes. In the UK, meeting summaries that name or describe identifiable people are personal data, and Article 16 of the UK GDPR gives individuals the right to have inaccurate personal data about them rectified, and incomplete data completed, including by way of a supplementary statement. The Information Commissioner's Office guidance says organisations should take reasonable steps to satisfy themselves that data is accurate when a rectification request is made, taking into account the individual's arguments and evidence. An employee who finds that a Copilot recap attributes to them a commitment they never made has, in principle, a route to have it corrected.
In practice that route depends on the employee noticing the error and having something to prove it with. If the transcript was never kept, the evidence is two sets of memories, one of which may have been rewritten by the summary itself. The right was designed for a world in which records and memories were separate things that could be compared.
Grievances and disciplinary hearings routinely rely on contemporaneous notes. An AI recap looks contemporaneous, with a timestamp and a corporate logo, but it carries no indication of which sentences reflect what was said and which are the model's reconstruction.
What the Law Currently Asks of Whom #
The European Union has come closest to allocating responsibility, though not for this specific harm.
Article 50 of the EU AI Act, whose core transparency obligations began applying on 2 August 2026, requires providers of AI systems that generate synthetic text, audio, images or video to ensure outputs are marked in a machine-readable format and detectable as artificially generated. It also requires deployers of AI systems that generate or manipulate text “published with the purpose of informing the public on matters of public interest” to disclose that the text was artificially generated. The Digital Omnibus, which pushed back the Act's high-risk obligations, left that date alone, although providers of generative systems already on the market before 2 August 2026 have until 2 December 2026 to meet the machine-readable marking requirement. The disclosure obligation falls away where the content “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication”.
That exception is revealing. It makes editorial responsibility, a named person or organisation standing behind the text, the alternative to disclosure. It does not make disclosure alone sufficient to discharge responsibility, and the new study suggests it should not. A label did not protect participants. A person accountable for checking the text before it reached them is a different kind of safeguard.
Liability is shifting too. The EU's revised Product Liability Directive, which member states must transpose by 9 December 2026, explicitly treats software, including AI systems, as a product for the purposes of no-fault liability, and eases the burden of proof for claimants dealing with technically complex systems. Its idea of damage centres on death, personal injury including medically recognised psychological harm, damage to property, and the destruction or corruption of data. A distorted memory of a meeting is unlikely to count. But the principle, that suppliers of software are answerable for defects in what they put on the market regardless of whether they were negligent, is a shift in the default assumption about who carries the risk.
In the United States, the picture is patchier. Beyond the Utah and California police laws, much of the current legal pressure on meeting assistants concerns consent rather than accuracy. In August 2025 a proposed class action, Brewer v. Otter.ai, was filed in federal court in California alleging that Otter's notetaker recorded conversations of people who had not consented, including non-users, and used the data to train its models. In October 2025 it was consolidated with similar suits as In re Otter.AI Privacy Litigation, and on 13 August 2026 Judge Eumi K. Lee refused to dismiss the core claims, allowing wiretap claims under federal and California law to proceed on the basis that Otter, by using recordings to develop its own models, could be treated as a third-party eavesdropper. The case illustrates where the legal system's attention currently sits: on whether the machine should have been listening, not on whether what it wrote afterwards was true.
Why the Reader Cannot Carry This Alone #
The easy answer to who should check the summary is the person reading it. That is what the disclaimers say, and what “human in the loop” means in most enterprise AI policies.
The study makes that answer hard to defend, for three reasons.
The first is that the checking mechanism is compromised by the thing being checked. Memory is the natural reference for testing a summary of an event you attended, but reading the summary changes the memory. Not every time, and not for everyone, but often enough that “check it against what you remember” is not a reliable safeguard.
The second is that the people most exposed to summaries are the people least equipped to check them. Summaries are used by people who did not attend, did not pay attention or did not read the original. That is the point of them. Asking those users to verify the summary against a source they were trying to avoid is asking them to do the work the product was meant to save.
The third is that confidence in the tool reduces checking. In a survey of 319 knowledge workers presented at the CHI conference in 2025, researchers at Microsoft Research and Carnegie Mellon University found that higher confidence in generative AI was associated with less critical thinking, while higher self-confidence was associated with more. The more reliable a tool seems, the less people scrutinise it, and a tool that is right most of the time trains exactly the complacency that lets its occasional errors through.
None of this means readers bear no responsibility. Attendees can take their own notes, readers can click through, officers can write their own account first. But a design that depends on users consistently doing the things the product was designed to spare them is not a design with a human safeguard. It is a design with a human scapegoat.
What Suppliers Could Actually Owe #
If the burden cannot sit only with readers, the obvious alternative is the organisations that build and deploy the summarisers. What would taking that burden seriously look like? It would start with measurement. The authors call for more technical audits of summarisation tools and for examining real-world uses such as body camera footage. Vendors selling summarisation for high-stakes settings could publish error rates on representative material, broken down by type: omissions of central events, substitutions, additions. A summariser that routinely omits the central event of a clip is not ready to summarise evidence.
It would continue with provenance. California's approach to police reports offers a model: keep the draft, keep the source, make it possible to see which words came from the machine. For meeting recaps, that could mean retaining source audio or transcripts for a defined period when a summary is created, or linking each sentence of a summary to the moment in the recording it draws on. Microsoft's recap already offers timeline markers and speaker navigation. Anchoring every claim to its evidence would make checking something a reader can do in seconds rather than a reconstruction from memory.
It would include friction in the right places. The study's authors suggest that noticing discrepancies helps, and that warnings delivered before reading might help. That points towards designs that ask users to recall something before showing them the summary, or that flag which parts of a recap are high-stakes (decisions, attributions, commitments) and invite confirmation from the people named. It points away from designs that deliver a finished, authoritative paragraph and a disclaimer.
And it would include restraint. The authors raise a question most product roadmaps do not: “a deeper question to consider is whether AI should be used in these contexts at all.” For some uses, such as drafting a police narrative from body camera audio before the officer has written their own account, the answer may be that the order of operations matters more than the quality of the model. Human recollection first, machine summary second, and a record of both.
Organisations that deploy summaries carry a share too. An employer that switches on recaps without transcripts and relies on them in a disciplinary process has made choices about evidence. A platform that pushes machine summaries of journalism to millions of lock screens has made choices about what its audience will remember. The EU's editorial responsibility test is a good instinct: somebody specific should be answerable for what the summary says.
Keeping Hold of the Afternoon You Lived #
There is a version of this story that is merely about quality control: the mistakes will get rarer and the problem will shrink. That may be partly true. But the study is not really about how often summaries are wrong. It is about what happens when they are. A human memory is not a hard drive that a summary sits beside. It is a reconstruction, rebuilt each time it is recalled, and fifty years of research have shown that the materials used in the rebuild include whatever we have been told since. For most of that history, the people doing the telling were other humans: lawyers asking leading questions, police officers suggesting details, newspapers and neighbours. They could be cross-examined. They could be blamed.
The summary is a new kind of teller, present at every meeting and every alert, speaking in a voice that sounds neutral and thorough. When it drops the collision from the crash, or puts a yield sign where a stop sign stood, it does not only produce a bad document. It produces people who remember a slightly different afternoon from the one they lived.
Right now, the burden of noticing falls on the people least able to notice, and responsibility for the error is spread so thinly that it lands on almost nobody. The disclaimers say the content may be inaccurate. The users are told to check. The checking, it turns out, runs through a memory the summary has already touched.
That arrangement looks unlikely to survive contact with a courtroom for long. The better move is to settle the question first: keep the source, mark the machine's words, measure the errors, and put a named person behind any summary that is going to be treated as the record. The past is hard enough to hold on to without software editing it for us.
References and Sources #
- Sim, Mattea, Yael Eiger and Tadayoshi Kohno. “AI-Enabled Human Memory Manipulation: Misleading AI-Generated Summaries Distort Human Memory.” arXiv:2609.28820v1, submitted 23 September 2026; to be presented at the Ninth AAAI/ACM Conference on AI, Ethics and Society (AIES), October 2026. https://arxiv.org/abs/2609.28820
- Soong, Kelyn. “New Research Finds That Misleading AI-Generated Summaries Can Distort Human Memory.” Georgetown University College of Arts and Sciences, 25 September 2026. https://college.georgetown.edu/news-story/new-research-finds-that-misleading-ai-generated-summaries-can-distort-human-memory/
- Neuroscience News. “Flawed AI Summaries Distort Eyewitness Memory.” 27 September 2026. https://neurosciencenews.com/ai-false-memory-llm-31261/
- The Independent. “AI summaries can warp your memory, study warns.” September 2026, as syndicated via inkl. https://www.inkl.com/news/ai-summaries-can-warp-your-memory-study-warns
- Loftus, Elizabeth F. and John C. Palmer. “Reconstruction of automobile destruction: An example of the interaction between language and memory.” Journal of Verbal Learning and Verbal Behavior, volume 13, issue 5, 1974, pages 585 to 589. https://www.sciencedirect.com/science/article/abs/pii/S0022537174800113
- Loftus, Elizabeth F., David G. Miller and Helen J. Burns. “Semantic integration of verbal information into a visual memory.” Journal of Experimental Psychology: Human Learning and Memory, volume 4, issue 1, 1978, pages 19 to 31. https://pubmed.ncbi.nlm.nih.gov/621467/
- Loftus, Elizabeth F. “Planting misinformation in the human mind: A 30-year investigation of the malleability of memory.” Learning and Memory, volume 12, 2005, pages 361 to 366. DOI 10.1101/lm.94705. https://learnmem.cshlp.org/content/12/4/361.full
- Chan, Samantha, Pat Pataranutaporn, Aditya Suri, Wazeer Zulfikar, Pattie Maes and Elizabeth F. Loftus. “Conversational AI Powered by Large Language Models Amplifies False Memories in Witness Interviews.” MIT Media Lab, arXiv:2408.04681, August 2024. https://arxiv.org/abs/2408.04681
- Microsoft Support. “Recap in Microsoft Teams.” Accessed 29 September 2026. https://support.microsoft.com/en-us/teams/meetings/recap-in-microsoft-teams
- Microsoft 365 Message Center. “MC1275312: Microsoft Teams: AI meeting recap without transcript to meet compliance policies.” Published 9 April 2026, archived at Merill Fernando's Message Center Archive. https://mc.merill.net/message/MC1275312
- Burke, Garance and Hilke Schellmann. “Researchers say an AI-powered transcription tool used in hospitals invents things no one ever said.” Associated Press, 26 October 2024; republished by Fortune and The Columbian. https://fortune.com/2024/10/26/openai-transcription-tool-whisper-hallucination-rate-ai-tools-hospitals-patients-doctors
- TechCrunch. “Apple s AI notification summaries for news after generating false alerts.” 16 January 2025; see also Axios, “Apple s AI-generated news alerts after fake headline notifications”, 17 January 2025. https://techcrunch.com/2025/01/16/apple-s-ai-notification-summaries-for-news-after-generating-false-alerts ; see also AppleInsider, “Apple brings back notification summaries for news in iOS 26”, 22 July 2025.https://appleinsider.com/articles/25/07/22/apple-brings-back-notification-summaries-for-news-in-ios-26
- BBC. “Representation of BBC News content in AI Assistants.” BBC Research, February 2025; as reported by The Register, “AI summaries turn real news into nonsense, BBC finds”, 12 February 2025. https://www.theregister.com/2025/02/12/bbc_ai_news_accuracy/
- European Broadcasting Union and BBC. “News Integrity in AI Assistants.” October 2025; and EBU press release, “Largest study of its kind shows AI assistants misrepresent news content 45% of the time, regardless of language or territory.” https://www.ebu.ch/news/2025/10/ai-s-systemic-distortion-of-news-is-consistent-across-languages-and-territories-international-study-by-public-service-broadcaste
- Reuters Institute for the Study of Journalism. “Digital News Report 2026: Overview and key findings.” University of Oxford, June 2026. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary
- Reid, Liz. Google Search blog post on AI Overviews, 30 May 2024; as reported by MIT Technology Review, “Why Google's AI Overviews gets things wrong”, 31 May 2024, and Forbes, 31 May 2024. https://www.technologyreview.com/2024/05/31/1093019/why-are-googles-ai-overviews-results-so-bad/
- Electronic Frontier Foundation. “Axon's Draft One Is Designed to Defy Transparency.” July 2025. https://www.eff.org/deeplinks/2025/07/axons-draft-one-designed-defy-transparency
- Guariglia, Matthew. “Victory! California Requires Transparency for AI Police Reports.” Electronic Frontier Foundation, 14 October 2025. https://www.eff.org/deeplinks/2025/10/victory-california-requires-transparency-ai-police-reports
- California Legislature. SB 524, Law enforcement agencies: artificial intelligence, 2025 to 2026 session; and Office of Senator Jesse Arreguín, “Governor Signs Bill Requiring Law Enforcement to Disclose Use of AI in Police Reports.” https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202520260SB524
- Utah Legislature. SB 180, Law Enforcement Usage of Artificial Intelligence, 2025 General Session; and Utah News Dispatch, “Deepfakes, policing, mental health and more: Lawmakers look to further regulate AI”, 26 March 2025. https://le.utah.gov/~2025/bills/static/SB0180.html
- Information Commissioner's Office. “Right to rectification.” UK GDPR guidance and resources. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/individual-rights/individual-rights/right-to-rectification/
- European Union. Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 50, Transparency obligations for providers and deployers of certain AI systems. https://artificialintelligenceact.eu/article/50/ ; see also Cooley, “EU AI Act: Transparency Obligations Take Effect 2 August 2026”, 3 August 2026.https://www.cooley.com/news/insight/2026/2026-08-03-eu-ai-act-transparency-obligations-take-effect-2-august-2026 ; and Goodwin, “Not Delayed, Not Deferred: EU AI Act Transparency Obligations Are Now in Force”, 3 August 2026.https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force
- European Union. Directive (EU) 2024/2853 on liability for defective products. Official Journal of the European Union, 18 November 2024. https://eur-lex.europa.eu/eli/dir/2024/2853/oj/eng
- NPR. “Class-action suit claims Otter AI secretly records private work conversations.” 15 August 2025. https://www.npr.org/2025/08/15/g-s1-83087/otter-ai-transcription-class-action-lawsuit ; and Baird Holm, “Core Privacy Claims Against AI Note Taking Platform Survive Dismissal” (In re Otter.AI Privacy Litigation, No. 5:25-cv-06911, N.D. Cal.), 24 August 2026.https://www.bairdholm.com/blog/core-privacy-claims-against-ai-note-taking-platform-survive-dismissal/
- Lee, Hao-Ping (Hank), and colleagues, Microsoft Research and Carnegie Mellon University. “The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.” Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, ACM. DOI 10.1145/3706598.3713778. https://dl.acm.org/doi/full/10.1145/3706598.3713778
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)
Listen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)