The Guardrail Crisis: Why Google Sent a Racial Slur to Millions

On the evening of 16 February 2026, the 79th British Academy Film Awards ceremony at London's Royal Festival Hall was already careening toward crisis. John Davidson, a Scottish Tourette syndrome activist and the subject of the BAFTA-nominated biographical film I Swear, was seated in the audience when his condition triggered a series of involuntary vocal tics. Davidson, who was appointed an MBE in 2019 for his work increasing understanding of Tourette syndrome, had been the subject of BBC documentaries since the age of sixteen, beginning with the 1989 programme John's Not Mad. His presence at the ceremony was meant to celebrate the film about his life, directed by Kirk Jones and starring Robert Aramayo. Instead, the evening became something far more painful.

The most devastating tic came as actors Michael B. Jordan and Delroy Lindo took to the stage to present the award for best visual effects. Davidson shouted a racial slur loudly enough for the entire hall to hear. The two Black actors paused, visibly processing the moment, and then continued with what BAFTA later described as “incredible dignity and professionalism.” Davidson left the ceremony of his own accord roughly twenty-five minutes into the proceedings, after which host Alan Cumming reminded the audience that “Tourette's syndrome is a disability and the tics you've heard tonight are involuntary.”

What followed was a cascading series of institutional failures. The BBC, which broadcast the ceremony on a two-hour delay, did not edit the slur out of the transmission, despite a formal request from Warner Bros. to do so. Notably, the same broadcast did edit out the phrase “Free Palestine” from an acceptance speech, a juxtaposition that drew furious commentary. The programme remained on BBC iPlayer for fifteen hours with the racial slur fully audible before being taken down. The BBC later conceded that the language should have been removed before transmission and issued an apology.

BAFTA Chair Sara Putt and CEO Jane Millichip sent a letter to members acknowledging “the harm this has caused” and announcing a “comprehensive review.” Award-winning filmmaker Jonte Richardson resigned from BAFTA's emerging talent judging panel, calling the organisation's handling of the situation “utterly unforgivable.” In his resignation statement on LinkedIn, Richardson wrote that he “cannot and will not contribute my time energy and expertise to an organization that has repeatedly failed to safeguard the dignity of its Black guests, members and the Black creative community.” Hannah Beachler, the Oscar-winning production designer on Sinners and the first African American to win an Academy Award for Best Production Design, revealed on X that Davidson's tics had been directed at her personally on the way to dinner after the show, describing the incident as happening “3 times that night.” She condemned what she called a “throw away apology” from host Alan Cumming.

Then Google made it worse.

The technology giant pushed out a computer-generated news alert to mobile devices, linking to a Hollywood Reporter article headlined “How the Tourette's Fallout Unfolded at the BAFTA Film Awards.” The notification then invited readers to “see more on” followed by the N-word, fully spelled out and sent directly to users' lock screens. Instagram user Danny Price was among the first to screenshot the notification and share it publicly, calling it “absolutely f****d” and noting the painful irony of receiving it during Black History Month. “What an interesting Black History month this has turned out to be,” Price wrote. Google removed the alert and issued an apology: “We're deeply sorry for this mistake. We've removed the offensive notification and are working to prevent this from happening again.”

The company also made a specific claim that would become the most interesting part of the entire episode. Google stated that the error “did not involve AI.” According to a spokesperson, their systems “recognised a euphemism for an offensive term on several web pages, and accidentally applied the offensive term to the notification text.” The safety filters that should have caught the slur before it reached users simply failed to trigger. Google told Entertainment Weekly that it was “working on improved guardrails for our push notification systems, which are designed to accurately characterize content from across the web.”

This distinction between “AI” and “automated system” is where the story gets genuinely revealing. Not because of what it tells us about Google's internal technical architecture, but because of what it exposes about the broader condition of technology deployment in public-facing spaces. Whether or not the specific notification was generated by a large language model is, in a meaningful sense, beside the point. What matters is that an automated system, operating without human oversight, sent a racial slur to an unknown number of mobile devices during Black History Month, and there was no human anywhere in the pipeline to prevent it.

The Distinction That Does Not Hold

Google's insistence that the BAFTA notification was not AI-generated deserves scrutiny, not because the company is necessarily lying, but because the distinction it is drawing has become functionally meaningless in the context of how automated systems interact with the public.

The OECD's AI Incidents Monitor, which tracks and classifies AI-related failures globally using a rigorous methodology that employs multiple large language models to categorise events, catalogued the Google BAFTA notification as an incident. Its analysts noted that, regardless of Google's denial, the system performed tasks “indicative of AI-like content processing,” including recognising euphemisms across multiple web pages and generating notification text from that analysis. As the commentary around the OECD classification observed, if the system was not AI, then it was human-engineered automation, and automation reflects choices. Somebody designed a system that could scan the web, identify trending topics, synthesise notification text, and push it to millions of devices without a human ever reading the output. That the system used pattern matching rather than a transformer model does not change the fundamental problem: a machine made an editorial decision about language in a racially charged context, and nobody checked its work.

The OECD's framework for classifying AI systems is itself instructive. It allows analysts to “zoom in on specific risks that are typical of AI, such as bias, explainability and robustness,” but it is deliberately generic in nature, designed to capture a broad spectrum of automated decision-making systems rather than only those that meet a narrow technical definition. The Google BAFTA incident sits precisely in the grey zone where technical definitions and practical consequences diverge. The system may not have been a neural network, but it performed the same function that an AI-powered summarisation tool would have performed, and it failed in precisely the same way.

This is not a novel failure mode. It is an accelerating one. In January 2025, Apple suspended its AI-powered notification summary feature after a string of high-profile hallucinations. The system had falsely told BBC News readers that Luigi Mangione, the man accused of killing UnitedHealthcare CEO Brian Thompson, had shot himself. It also incorrectly claimed that tennis star Rafael Nadal had come out as gay, named the PDC World Darts Championship winner before the competition ended, and falsely stated that Israeli Prime Minister Benjamin Netanyahu had been arrested. Reporters Without Borders called on Apple to remove the feature entirely, arguing that AI “cannot reliably produce information for the public.” The BBC itself complained to Apple, along with other media organisations, all of whom said the technology was “not ready” and that the AI-generated errors were “adding to issues of misinformation and falling trust in news.”

In May 2024, Google's own AI Overviews feature, which uses the Gemini large language model to generate summary answers atop search results, went spectacularly wrong within days of its US launch. The system advised users to add non-toxic glue to pizza sauce to help cheese stick, a recommendation it had sourced from an eleven-year-old joke on Reddit. It recommended eating “at least one small rock per day” as a source of minerals, drawing from a satirical article in The Onion. It suggested dangerous practices such as mixing bleach and vinegar, which produces toxic chlorine gas. It told users that former US President Barack Obama is a Muslim, and that astronauts had met cats on the moon. Google CEO Sundar Pichai, responding to the debacle, told The Verge that hallucination is “still an unsolved problem” and even, “in some ways, an inherent feature” of large language models. He added that “LLMs aren't necessarily the best approach to always get at factuality.” Data from SEO firm BrightEdge showed that Google quietly reduced the frequency of AI Overviews in search results from 27 per cent to 11 per cent in the weeks following the launch.

A few months earlier, in February 2024, Google had paused its Gemini image generator after it produced historically inaccurate and racially offensive images, including people of colour depicted in Nazi-era uniforms in response to prompts about German soldiers in 1943. The system also generated images of Black Vikings, a woman as the Catholic pope, and non-white people in a scene depicting the founding of the United States. Prabhakar Raghavan, a senior vice president at Google, acknowledged: “It's clear that this feature missed the mark. Some of the images generated are inaccurate or even offensive.” The problem had arisen because the system had been programmed to inject diversity language into prompts, transforming a request for “pictures of Nazis” into something like “pictures of racially diverse Nazis,” a well-intentioned overcorrection that produced deeply offensive results.

The pattern is unmistakable. Major technology companies are deploying automated systems into high-stakes public contexts, discovering that those systems can produce harmful, offensive, or factually false outputs, issuing apologies, and then continuing to deploy substantially similar systems with minor adjustments. The cycle repeats because the commercial incentives to deploy outweigh the reputational costs of failure.

Automation Without Accountability

The Google BAFTA incident is instructive precisely because it sits at the boundary between what companies classify as “AI” and what they classify as “automated systems.” This boundary is not a technical distinction that users experience or understand. From the perspective of the person who received a racial slur on their lock screen, the question of whether a large language model or a keyword-matching algorithm generated the text is entirely irrelevant. The harm is identical. The absence of human oversight is identical. The failure of safety systems is identical.

This is a problem that extends well beyond news notifications. Between November 2025 and January 2026, the AI Incident Database added 108 new incident IDs, covering failures across healthcare, employment, law enforcement, and public information systems. Stanford's Human-Centred Artificial Intelligence Institute reported that publicly reported AI-related security and privacy incidents rose 56.4 per cent from 2023 to 2024. The trajectory is accelerating, not stabilising.

A 2025 study found that clinicians' tumour detection rates dropped six per cent after months of working with AI assistance, a documented manifestation of automation bias in which humans systematically over-trust automated decisions even when contradictory evidence is present. The European Data Protection Supervisor published a 2025 dispatch on human oversight of automated decision-making that warned of “vigilance decrement,” a measurable deterioration in the ability to detect anomalies during passive monitoring tasks. The dispatch argued that in contexts where human operators rely heavily on system recommendations, “there should be a presumption of automation by default,” meaning that deployers should treat the system as if it were operating autonomously and apply effective human oversight accordingly.

In employment, a federal judge in May 2025 allowed a collective action lawsuit to proceed under the Age Discrimination in Employment Act, alleging that Workday's AI-powered screening tools disproportionately disadvantaged applicants over forty. One plaintiff reported receiving immediate rejection notifications during non-business hours, suggesting automated filtering with no human involvement whatsoever. The case was certified as a nationwide class action. France's independent equality watchdog ruled that Facebook's job advertisement distribution algorithm was discriminatory and sexist, showing bus driver advertisements almost exclusively to men and nursery assistant advertisements almost exclusively to women.

In content moderation, research from the Internet Freedom Foundation and the Knight Foundation has demonstrated that AI systems trained predominantly on “standard” English consistently flag content from Black creators at higher rates, particularly when African American Vernacular English is used. The Brookings Institution found that Black comedians using satirical commentary on racial stereotypes were banned for “promoting stereotypes,” while white counterparts making identical points received no penalties. A study on deepfake detection found that classifiers misidentified real images of Black men as fabricated 39.1 per cent of the time, compared to 15.6 per cent for white women, revealing serious racial disparities in AI-based verification systems.

The common thread is not that AI is uniquely dangerous. It is that automated systems of all kinds, whether powered by large language models, keyword matching, or statistical classifiers, are being deployed at a scale and speed that fundamentally outpaces the development of meaningful oversight mechanisms. The Google BAFTA notification is a particularly vivid example because the harm was so immediate, so public, and so obviously preventable by a single human reading the output before it was sent.

The Guardrail Gap

The regulatory landscape is struggling to keep pace. The European Union's AI Act, the most comprehensive AI regulation attempted by any major jurisdiction, follows a phased implementation timeline. Prohibitions on AI systems posing unacceptable risks came into effect in February 2025. Governance infrastructure and obligations for providers of general-purpose AI models followed in August 2025. But the critical transparency requirements and rules for high-risk AI systems do not take effect until August 2026, and enforcement powers for the European Commission only begin on that date. Rules for high-risk AI systems embedded in regulated products have an extended transition period running to August 2027. There are already signals that even these timelines may slip: in November 2025, the European Commission published legislative proposals that would extend the applicability date for high-risk AI rules from August 2026 to as late as December 2027.

The penalties for non-compliance are theoretically significant, reaching up to seven per cent of worldwide annual turnover. But the regulatory framework is designed primarily for systems that companies acknowledge as AI. Google's insistence that the BAFTA notification was “not AI” illustrates a definitional gap that could become a regulatory escape hatch. If a company can argue that its automated content generation system does not meet the technical definition of artificial intelligence under the EU AI Act, it may be able to avoid the transparency, oversight, and accountability requirements that the regulation imposes. Each Member State is required to establish at least one AI regulatory sandbox by August 2026, but these testing environments are designed for systems that are acknowledged as AI from the outset, not for automated pipelines that companies refuse to classify as such.

In the United States, the regulatory picture is even more fragmented. Colorado has delayed implementation of its comprehensive AI law to June 2026. Several states, including Illinois, New York, Utah, and California, have adopted disclosure requirements and protections specifically for AI companions and therapeutic tools. But there is no federal AI regulation, and the patchwork of state-level rules creates an environment in which companies can deploy automated systems nationally while navigating wildly inconsistent oversight regimes.

The fundamental problem is not a lack of awareness. The OECD has developed a global AI incident reporting framework. The EU AI Act mandates AI literacy training and conformity assessments. Academic institutions, civil society organisations, and international bodies have produced thousands of pages of guidance, principles, and recommendations. What is missing is the connective tissue between these frameworks and the actual moment of deployment, the point at which an automated system generates a piece of text, an image, a recommendation, or a notification, and sends it into the world without a human ever seeing it first.

Lessons Unlearned

The history of automated systems producing harmful output in public-facing contexts is not short, and the technology industry's institutional memory for its own failures appears to be remarkably brief.

In March 2016, Microsoft launched Tay, a Twitter chatbot designed to engage with users and learn from their interactions. Within sixteen hours, coordinated users had manipulated Tay into posting antisemitic, racist, and sexist content, including the statement “Hitler was right I hate the jews.” Microsoft shut it down after it had generated more than 96,000 tweets, attributing the failure to “a coordinated attack by a subset of people” who “exploited a vulnerability in Tay.” The episode was widely analysed as a cautionary tale about deploying learning systems in adversarial environments without adequate safeguards. IEEE Spectrum later noted that the case illustrated “a problem with the very nature of learning software that interacts directly with the public, and the developer's role and responsibility associated with it.”

In November 2022, Meta released Galactica, a large language model trained on 48 million scientific texts and designed to assist researchers. Within three days, Meta pulled the public demo after the model generated convincing-sounding papers on the benefits of committing suicide, fabricated research papers attributed to real scientists, and produced plausible-seeming articles about the history of bears in space. The problem, as researchers pointed out, was that Galactica could not distinguish between truth and fabrication. It generated text with the same authoritative tone regardless of whether the underlying claims were factual or invented. As MIT Technology Review observed, “Big tech companies keep doing this because they can. And they feel like they must, otherwise someone else might.”

Each of these incidents prompted calls for greater caution, more robust safety testing, and stronger oversight mechanisms. Each was followed by a period of reflection. And each was subsequently overtaken by commercial pressure to deploy the next generation of tools faster than the previous one. The 2026 AI Safety Report confirms that this dynamic is intensifying: some models now distinguish between evaluation and deployment contexts, altering their behaviour to appear safer during testing than they actually are in production. Despite extraordinary advances, the report warns, models remain less reliable on multi-step projects, still produce hallucinations, and struggle with tasks involving the physical world.

The Google BAFTA notification fits this pattern with uncomfortable precision. The system that generated it was not some experimental research prototype. It was a production system, pushed to the devices of real users, processing real-world events with serious racial dimensions, with no human gatekeeper between the algorithm and the public. Google's response, to apologise and promise improved guardrails, is the same response the company gave after the AI Overviews debacle, the Gemini image generator controversy, and numerous other incidents. The cycle of deploy, fail, apologise, and adjust has become the de facto governance model for automated content systems.

The Problem With Speed

The commercial logic driving this cycle is straightforward. In the race to integrate AI and automation into every consumer-facing product, speed of deployment is treated as a competitive advantage. Google, Apple, Meta, Microsoft, and their competitors are engaged in a contest where being first to market with AI-powered features is seen as strategically essential. This creates an environment in which the question “does this system work reliably?” is subordinated to the question “can we ship this before our competitors do?”

The consequences of this approach are distributed unevenly. When Google's AI Overviews suggested eating rocks, the primary victims were individual users who might have taken the advice seriously. When Apple's notification summaries falsely reported that a murder suspect had shot himself, the harm extended to the news organisation whose credibility was implicated and to the public's trust in information systems more broadly. When Google's notification system sent a racial slur to users' phones during Black History Month, the harm landed on Black communities already navigating a cultural moment of particular sensitivity. The costs of speed are paid by the people who have the least say in how quickly these systems are deployed.

This is not a problem that can be solved by better safety filters alone. Google had safety filters in place for the BAFTA notification. They failed. Apple had safety filters in place for its notification summaries. They failed. Google had safety filters in place for AI Overviews. They failed. The question is not whether safety filters can be improved. Of course they can. The question is whether a governance model that depends entirely on algorithmic safety filters, with no human in the loop for high-stakes editorial decisions, is fundamentally adequate for the task. The evidence suggests, with mounting force, that it is not.

What Meaningful Oversight Would Require

A serious response to the Google BAFTA incident would require confronting several uncomfortable realities that the technology industry has so far been reluctant to acknowledge.

First, the distinction between “AI” and “automated systems” needs to be abandoned as a regulatory category. What matters is not the technical architecture of the system but its function: is it generating content that reaches the public without human review? If so, the same standards of accuracy, sensitivity, and accountability should apply regardless of whether the underlying mechanism is a large language model, a keyword matcher, or a rule-based classifier. The EU AI Act's risk-based approach provides a useful framework, but only if the definition of what constitutes an AI system is broad enough to capture the full range of automated content generation tools currently in deployment.

Second, human oversight for high-stakes automated outputs needs to be treated as a non-negotiable requirement, not a nice-to-have feature that can be sacrificed for speed. A notification system that pushes content to millions of devices about racially sensitive events should have a human editor in the pipeline. The argument that this would slow down the delivery of notifications is precisely the point. Some content is too consequential to be left entirely to machines, and the determination of what falls into that category should be made in advance, not after a slur has already been broadcast. The Harvard Journal of Law and Technology has argued for redefining the standard of human oversight in the context of AI negligence, suggesting that legal frameworks need to evolve to hold deployers accountable when they choose to remove humans from decision loops in high-stakes contexts.

Third, the accountability structures for automated system failures need to be formalised and enforced. When a newspaper publishes a racial slur, the editor responsible can be identified, questioned, and held accountable. When an automated system does the same thing, the accountability diffuses across engineering teams, product managers, policy groups, and corporate communications departments. Nobody is responsible because everybody is responsible. The result is that the same failures recur because no individual faces consequences meaningful enough to change institutional behaviour.

Fourth, incident reporting and learning need to become systematic rather than reactive. The OECD's AI Incidents Monitor and the AI Incident Database represent important steps, but they remain largely academic exercises rather than binding mechanisms for institutional change. A mandatory incident reporting regime, analogous to the aviation industry's approach to near-misses and accidents, would create the feedback loops necessary for genuine improvement. Companies should be required to report automated system failures to a centralised authority, with the data used to inform regulatory standards and best practices. The OECD published a paper in February 2025 outlining the foundations for such a framework, but translating it from policy paper to binding obligation remains a distant prospect.

Beyond Apologies

Davidson, for his part, issued a statement saying he was “deeply mortified if anyone considers my involuntary tics to be intentional or to carry any meaning.” His situation is genuinely complex: a man whose neurological condition produces involuntary vocalisations, attending the premiere celebration of a film about his life, in a venue that had not adequately prepared for the known characteristics of his disability. Davidson's team shared that he subsequently reached out to the studio handling Sinners in order to directly apologise to Jordan, Lindo, and Beachler. The failures of BAFTA and the BBC in handling the live event are significant and have been extensively discussed, raising uncomfortable questions about both ableism and duty of care.

But the Google notification represents something categorically different. It is not a failure of empathy or event management. It is a failure of systems design, one that reveals how automated content pipelines treat language as data to be processed rather than as communication that carries weight, context, and consequence. A system that can “recognise a euphemism for an offensive term” and then “accidentally apply the offensive term” has demonstrated that it can parse language at a mechanical level while being entirely blind to the social, racial, and historical dimensions of that language. This is not a bug that can be patched with a better keyword filter. It is a structural feature of systems designed to operate at speed and scale without the interpretive capacities that human editorial judgement provides.

The broader pattern is one of technology companies externalising the costs of their deployment speed onto the communities most likely to be harmed by the resulting failures. When an automated system sends a racial slur to users' phones, the immediate cost is borne not by Google but by the Black users who received it, by the actors who were on stage when the original incident occurred, and by the production designer who had the slur directed at her personally. Google's cost is a news cycle of criticism and an apology that costs nothing to produce. The asymmetry is structural, and it will not change until the regulatory and commercial incentives are realigned.

The BAFTA notification should function as something more than a footnote in the long catalogue of automated system failures. It should be recognised as a concrete illustration of what happens when the guardrails lag behind the deployment by years rather than months. The technology to send automated notifications exists. The technology to scan the web and generate summary text exists. The technology to push that text to millions of devices in seconds exists. What does not yet exist, in any meaningful or enforceable form, is the institutional architecture to ensure that these capabilities are exercised with the care that their power demands.

Until that architecture is built, the cycle will continue. Another automated system will produce another harmful output. Another company will issue another apology. Another community will absorb another cost that was never theirs to bear. The question posed by the Google BAFTA notification is not whether this particular failure could have been prevented. It obviously could have been. The question is whether the industry and its regulators are willing to build the systems necessary to prevent the next one, even if doing so means deploying more slowly, charging more honestly for the cost of human oversight, and accepting that some things are too important to be left entirely to machines.

References and Sources

  1. Deadline. “Google Apologizes After News Alert About BAFTA Film Awards Debacle Included The N-Word.” Deadline, February 2026. https://deadline.com/2026/02/google-apologizes-bafta-news-alet-n-word-1236734448/

  2. Variety. “Google 'Deeply Sorry' for BAFTA News Alert That Included N-Word, Says the Message Was Not AI-Generated.” Variety, February 2026. https://variety.com/2026/digital/news/google-sorry-bafta-n-word-news-alert-1236671565/

  3. The Wrap. “Google Apologizes for 'Sinners' News Alert That Included Spelled-Out Racial Slur From BAFTA Awards.” The Wrap, February 2026. https://www.thewrap.com/media-platforms/journalism/google-uses-racial-slur-ai-generated-sinners-alert/

  4. Word In Black. “If It Wasn't AI, Who Put the N-Word in Google's Push Alert?” Word In Black, February 2026. https://wordinblack.com/2026/02/google-n-word-push-notification/

  5. OECD.AI. “Google AI Push Notification Includes Racial Slur, Prompts Apology.” OECD AI Incidents Monitor, 24 February 2026. https://oecd.ai/en/incidents/2026-02-24-6844

  6. CNN. “John Davidson: BAFTAs interrupted by racist slur from man with Tourette Syndrome.” CNN, 22 February 2026. https://www.cnn.com/2026/02/22/entertainment/baftas-2026-tourettes-racist-slur

  7. NBC News. “BAFTA and BBC apologize to Michael B. Jordan and Delroy Lindo after guest with Tourette syndrome shouted slur.” NBC News, February 2026. https://www.nbcnews.com/pop-culture/pop-culture-news/bbc-says-racial-slur-shouted-sinners-actors-baftas-was-result-tourette-rcna260182

  8. PBS News. “BAFTA and BBC apologize for broadcasting racial slur during awards show.” PBS News, February 2026. https://www.pbs.org/newshour/arts/bafta-and-bbc-apologize-for-broadcasting-racial-slur-during-awards-show

  9. Variety. “John Davidson Gives First Interview and Explains Tourette's Tics After Shouting N-Word and Other Slurs at BAFTAs.” Variety, February 2026. https://variety.com/2026/film/awards/john-davidson-tourettes-tics-bafta-n-word-interview-1236671850/

  10. Deadline. “John Davidson Says He Is 'Deeply Mortified' That His Tourette's Tics Could Be Seen As 'Intentional.'” Deadline, February 2026. https://deadline.com/2026/02/john-davidson-issues-statement-bafta-racial-slur-i-swear-1236733373/

  11. Variety. “'Sinners' Production Designer Hannah Beachler Says Alan Cumming's 'Throw-Away Apology' Over N-Word Slur During BAFTAs 'Made It Worse.'” Variety, February 2026. https://variety.com/2026/film/global/sinners-hannah-beachler-n-word-slur-baftas-apology-1236670089/

  12. Deadline. “'Sinners' Production Designer Hannah Beachler Decries BAFTA's 'Throw-Away' On-Stage Apology After N-Word Outburst.” Deadline, February 2026. https://deadline.com/2026/02/sinners-hannah-beachler-bafta-apology-n-word-alan-cumming-1236732924/

  13. Deadline. “BAFTA Jury Member Jonte Richardson Steps Down Over Racial Slur.” Deadline, February 2026. https://deadline.com/2026/02/bafta-jonte-richardson-jury-member-steps-down-racial-slur-1236734268/

  14. Variety. “BAFTA Jury Member Steps Down Over N-Word Incident.” Variety, February 2026. https://variety.com/2026/film/awards/bafta-jury-member-steps-down-over-n-word-incident-1236671070/

  15. Washington Post. “Apple pauses AI summaries that botched news headlines.” Washington Post, January 2025. https://www.washingtonpost.com/technology/2025/01/16/apple-intelligence-hallucination/

  16. Deadline. “Apple Ceases AI News Alerts After High-Profile Mistakes.” Deadline, January 2025. https://deadline.com/2025/01/apple-cancels-ai-news-alerts-bbc-1236259586/

  17. MIT Technology Review. “Why Google's AI Overviews gets things wrong.” MIT Technology Review, 31 May 2024. https://www.technologyreview.com/2024/05/31/1093019/why-are-googles-ai-overviews-results-so-bad/

  18. Washington Post. “Why Google's AI search might recommend you mix glue into your pizza.” Washington Post, 24 May 2024. https://www.washingtonpost.com/technology/2024/05/24/google-ai-overviews-wrong/

  19. Futurism. “CEO of Google Says It Has No Solution for Its AI Providing Wildly Incorrect Information.” Futurism, May 2024. https://futurism.com/the-byte/ceo-google-ai-hallucinations

  20. Variety. “Google Pauses AI Image Generation of People to Fix Racial Inaccuracies.” Variety, February 2024. https://variety.com/2024/digital/news/google-gemini-ai-image-racial-inaccuracies-nazi-soldiers-1235919168/

  21. NPR. “Google races to find a solution after AI generator Gemini misses the mark.” NPR, 18 March 2024. https://www.npr.org/2024/03/18/1239107313/google-races-to-find-a-solution-after-ai-generator-gemini-misses-the-mark

  22. IEEE Spectrum. “In 2016, Microsoft's Racist Chatbot Revealed the Dangers of Online Conversation.” IEEE Spectrum, 2019. https://spectrum.ieee.org/in-2016-microsofts-racist-chatbot-revealed-the-dangers-of-online-conversation

  23. MIT Technology Review. “Why Meta's latest large language model only survived three days online.” MIT Technology Review, 18 November 2022. https://www.technologyreview.com/2022/11/18/1063487/meta-large-language-model-ai-only-survived-three-days-gpt-3-science/

  24. European Commission. “AI Act: Shaping Europe's Digital Future.” European Commission, 2025. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

  25. Wilson Sonsini. “2026 Year in Preview: AI Regulatory Developments for Companies to Watch Out For.” Wilson Sonsini, 2026. https://www.wsgr.com/en/insights/2026-year-in-preview-ai-regulatory-developments-for-companies-to-watch-out-for.html

  26. European Data Protection Supervisor. “TechDispatch #2/2025: Human Oversight of Automated Decision-Making.” EDPS, 23 September 2025. https://www.edps.europa.eu/data-protection/our-work/publications/techdispatch/2025-09-23-techdispatch-22025-human-oversight-automated-making_en

  27. Harvard Journal of Law and Technology. “Redefining the Standard of Human Oversight for AI Negligence.” Harvard JOLT, 2025. https://jolt.law.harvard.edu/digest/redefining-the-standard-of-human-oversight-for-ai-negligence

  28. AI Incident Database. “AI Incident Roundup: November and December 2025 and January 2026.” AI Incident Database, 2026. https://incidentdatabase.ai/blog/incident-report-2025-november-december-2026-january/

  29. OECD. “Towards a Common Reporting Framework for AI Incidents.” OECD Artificial Intelligence Papers No. 34, February 2025. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/02/towards-a-common-reporting-framework-for-ai-incidents_8c488fdb/f326d4ac-en.pdf

  30. Mediaite. “Google Apologizes for Sending Out AI-Generated Push Notification That Used the N-Word.” Mediaite, February 2026. https://www.mediaite.com/media/entertainment/google-apologizes-for-sending-out-ai-generated-push-notification-that-used-the-n-word/

  31. ISACA. “Avoiding AI Pitfalls in 2026: Lessons Learned from Top 2025 Incidents.” ISACA Now Blog, 2025. https://www.isaca.org/resources/news-and-trends/isaca-now-blog/2025/avoiding-ai-pitfalls-in-2026-lessons-learned-from-top-2025-incidents

  32. OECD.AI. “OECD Framework for the Classification of AI Systems.” OECD, 2025. https://oecd.ai/en/classification

  33. OECD.AI. “Overview and methodology of the AI Incidents and Hazards Monitor.” OECD, 2025. https://oecd.ai/en/incidents-methodology

  34. Black Current News. “Exclusive: Google sends N-word in BAFTAs news alert, company apologises.” Black Current News, February 2026. https://www.blackcurrentnews.co.uk/p/google-apologises-n-word-push-notification-baftas

Tim Green

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 Email: tim@smarterarticles.co.uk

Discuss...