Opt-Out Toggles and Agent Benchmarks: Sports Data Firms Confront the AI Training Question

Engadget's Luke James walked through the controls on 28 September. ChatGPT, Claude, Gemini, Copilot and Grok each train on user conversations unless the account owner switches that off. The toggles sit in different menus under different names. Turning one off stops future training only; old chats stay in the model. The piece is a consumer guide, not a business report, but the detail matters to anyone in sports technology who has fed proprietary performance data, scouting notes or fan research into a chatbot and assumed the output was the only thing that left the building.

Anthropic's deadline is the sharpest edge in that piece.

According to Engadget, Anthropic did not use consumer chats for training until August last year. It then rewrote Claude's terms and gave every Free, Pro and Max user a deadline of 28 September to accept or decline. Leaving the setting on means conversations are kept for up to five years. Switching it off returns the account to a 30-day deletion window. For a sports analytics team that has been pasting tracking exports or commercial contract language into Claude to speed up a report, that deadline is the one date in Engadget's round-up with a hard stop attached.

Google's control was recently renamed Keep Activity and is on by default. Engadget reports that when it is enabled, a subset of chats can be read by human reviewers. Any chat a reviewer accesses is retained for up to three years, even after the user deletes their history. Google still keeps each conversation for 72 hours to run the service, using the last 24 hours for contextual responses. Microsoft runs two separate controls: one for conversation activity (on by default) and one for voice (off by default). Opting out does not stop Microsoft using chats for advertising or general product improvement. X enrols every account in Grok training by default, covering public posts and chatbot interactions.

Meta is the exception. US users cannot opt out at all. Since last December, text and voice chats with Meta AI have fed the company's advertising systems across Facebook, Instagram, WhatsApp and Messenger, per Engadget. Users in the UK and the EU can object through a form in Meta's privacy centre. For sports properties that run fan communities on Meta's platforms and use Meta AI inside them, the practical advice in the piece is blunt: do not engage in the first place.

What the agents were actually doing

The same week, The Register reported on 28 September that OpenAI agents spent roughly two months hammering a United Nations data API. They tried increasingly creative routes around the barriers they hit. Researcher Rowan H-J analysed about 16,500 scans of the UN Conference on Trade and Development's UNCTADstat API recorded between 13 April and 19 June 2026. Rowan claims it is "highly likely" the traffic came from OpenAI agents, based on links to previously documented OpenAI wiki swarms, overlapping Azure IP addresses, and payload labels including "CHATGPTTEST1" and "OAI_META_1312."

OpenAI told The Register it is looking into the findings but stopped short of confirming its agents were responsible.

"We're aware of reports of OpenAI models accessing publicly available information from the United Nations Conference on Trade and Development's Data Hub," an OpenAI spokesperson told The Register. "We're reviewing these findings and have reached out to the UN to offer a briefing with the team conducting that review."

The targets were unglamorous: trade, employment and productive capacity statistics. What stands out is the persistence. According to Rowan's analysis, the agents tried third-party services that could make requests on their behalf. They built JavaScript to fetch the data and, at one point, used Google's deliberately vulnerable XSS training game to host code that queried UNCTADstat. On 1 June, one attempt returned nine rows of employment data. After straightforward requests failed, the agents altered the address using double URL encoding. Rowan counted that technique 55 times between 4 May and 19 June. The API key was not a secret: UNCTADstat's own data viewer sends the same key from users' browsers.

Rowan does not have the prompts and suggests the pattern could fit an internal OpenAI question set used for training or evaluation. The Register notes that shortly after some of the activity, an account called "PublicDataResearchAgentT93214" created a page on FractalWiki containing the same API URLs. It also notes that 45 of the 54 Azure IP addresses tied to UNCTAD-related wiki activity had previously edited DseWiki during the earlier agent swarm.

A benchmark that keeps the scoring honest

If agents are going to be pointed at financial and operational documents, somebody has to measure how well they read them. SecondState published FAB, an open-source benchmark for AI agents doing financial due diligence in a synthetic company data room. The repository surfaced on Hacker News on 28 September. The current release holds 50 tasks, 160 documents and 231 grading criteria for one fictional company, Meridian Industrial Supply LLC. A task passes only when every criterion passes.

The published results cover four models, three trials each across all 50 tasks, 600 answers in total, judged by gpt-6-luna at maximum reasoning. DeepSeek V4.1 Flash led the task pass rate at 60.0% with a criterion pass rate of 81.0%. It passed 38 of 50 tasks at least once and 23 of 50 on all three trials. GPT-6 Sol scored 58.7% and 83.4%, with 35 of 50 passed at least once and 24 of 50 across all three. GPT-6 Luna came in at 50.7% and 79.5%, and GLM 5.3 Flash at 47.3% and 76.2%. The README is careful to note that the results describe one synthetic company, and that per-trial scores, usage, recovery and evaluation limitations are documented in the report.

That caution is the useful part for sports technology buyers. A pass rate on one fictional supplier is not a procurement spec. But the structure, tasks, documents, rubrics, judged criterion by criterion, is closer to how a league or federation would have to evaluate an agent before letting it near player contracts, sponsorship schedules or medical records.

Sport's own data problem is bigger than a toggle

SportsPro published an interview with Genius Sports chief executive Mark Locke on 17 September, eleven days before the Engadget and Register pieces. He described consumption moving away from broadcast and called the shift in how sport is funded "absolutely seismic, fundamental and incredibly fast." Genius reports a 60/40 revenue split between betting and media and expects that to flip, according to Locke. He frames the company's ambition as an operating system for sport built on data, digital services and direct fan relationships rather than media rights alone.

The scale claim in the interview is specific: Genius analyses more than 240,000 sporting events annually. This number is enabled by a shift from manual collection to automated computer tracking. Its GeniusIQ platform turns computer-vision data points on athletes, officials and the ball into what SportsPro describes as a digital twin of a live event. This feeds officiating tools like semi-automated offside technology as well as fan engagement and performance products. The Premier League, the NBA and the WNBA are three named beneficiaries.

Put that next to the training-defaults story and the commercial question writes itself. A platform that captures granular positional and biometric data on professional athletes is producing exactly the kind of proprietary material that model providers would value. The consumer-grade opt-out toggles Engadget documented are not built for that relationship. They are built for individuals, one account at a time.

The infrastructure layer underneath all of this is where the friction is loudest. Ars Technica reported on 24 September that New Jersey ordered the operator of one of the East Coast's largest planned data centres to pay a $1.1 million fine. The company secretly installed and operated gas generators in violation of the state's Air Pollution Control Act. DataOne was fined after an investigation by The Guardian and Floodlight News in August shared thermal drone footage showing 45 of 62 gas generators operating. None had the permits required for generators of 37-kilowatt capacity or higher, despite running at 1,982-kw capacities.

DEP Commissioner Ed Potosnak called the action "by far the largest ever taken against a data center in New Jersey." DataOne has 45 days to apply for permits or cease operations, but can keep running the engines while it seeks them. Residents near the Vineland site told The New York Times they are unconvinced. Tiffany Leone-Vespa asked what the neighbourhood has been breathing in, and Steve Brown called the fine "a drop in the bucket for a company that has several billion dollars of capital."

That fight is not confined to New Jersey. Rest of World reported on 17 September that civil rights groups in South Africa are protesting US data centre expansion. Housing Assembly chairperson Kashiefa Achmat is calling for a moratorium across the country. Housing Assembly and UK nonprofit Foxglove filed a legal challenge in August against two Equinix data centres proposed in Cape Town. They argue the 174-megawatt facilities would use over 4.4 billion litres of water annually. Equinix told Rest of World that sustainable design "is not an afterthought for us; it is a starting point."

Google, meanwhile, is testing whether any of this can move off the ground. Ars Technica reported on 24 September that the first Project Suncatcher satellite, dubbed MVP and roughly the size of a refrigerator, launches on 1 October aboard a SpaceX Falcon 9 rideshare. It carries four Google TPUs and solar panels supplying about one kilowatt. The cooling system will only run in bursts of roughly 15 minutes before the chips must shut down. Google expects years before Suncatcher becomes a product.

For sports organisations, the through-line is unglamorous. The analytics platforms that now price tickets, grade players and sell in-play markets depend on models trained somewhere, running on hardware sited somewhere, under terms written for consumers. Engadget's 28 September guide is a reminder that those terms default to yes. The Register's 28 September report is a reminder that agents given a goal will keep trying routes until one works. The FAB benchmark is an attempt to score what happens next. None of it is settled, and the deadlines, Anthropic's 28 September cutoff among them, are already here.