Chat Picker

2026 AI Tool Regulation: Compliance and User Rights Protection

An afternoon test for checking AI plan terms and user controls against EU rules listed for August 2026 and separate UK guidance.

Sources checked 3 Oct 2026

For 2026 AI tool regulation, vendor pages settle account terms, not legal compliance: OpenAI’s data controls, Anthropic’s model-training page, and Gemini’s Privacy Hub describe different privacy paths. They were as read on October 1, 2026. Chat Picker has not tested ChatGPT, Claude, or Gemini for compliance, accuracy, or user-rights handling; check official duties first, then your own account.

Know these limits before you start

As read on October 1, 2026, the European Commission’s AI Act page defines four risk levels and says systems considered a clear threat to safety, livelihoods, or rights are banned. It lists individual criminal-offence risk assessment or prediction as one of nine prohibited practices. The first eight became effective in February 2025; Prohibition 9 is listed for December 2026. Transparency rules are listed for August 2026, and strict pre-market duties for high-risk systems from December 2, 2027. Employment is among the listed high-risk uses. The page also says chatbot users should know they are interacting with a machine, generated content should be identifiable, and deepfakes and public-interest text should be clearly labeled.

The same page says minimal- or no-risk systems face no rules and that most AI systems currently used in the EU fall into those categories. The Commission’s GPAI guidance says provider obligations entered application on August 2, 2025, and enforcement powers apply from August 2, 2026; older models must comply by August 2, 2027. The guidance is not legally binding, but it reflects the Commission’s interpretation. It says providers of the most advanced models that pose systemic risks must notify the AI Office.

The UK ICO guidance is separate. The ICO says its March 15, 2023 update is under review because of changes made by the Data (Use and Access) Act and may change. It addresses transparency and explainability, plus accuracy and fairness across the AI lifecycle. Do not merge it with the EU timetable.

Product controls are not a compliance switch. OpenAI’s data controls say availability depends on sign-in, plan, and workspace; disabling model improvement does not delete saved chats, and temporary chats stay out of history and memory. Signed-in Free, Go, Plus, and Pro users can request a data copy and delete their account. Anthropic’s training page says Claude chats and coding sessions improve models when the user allows it, incognito chats are excluded, and safety-flagged conversations may be used for safeguards. Gemini personalization requires a personal Google Account, can use past chats and connected app content, and is not available to everyone. OpenAI and Claude warn that answers can be incorrect or misleading; Gemini documents a Sources button when sources are available, not an accuracy guarantee.

The test plan

  1. Classify the use. Prepare a fictional retailer case involving applicant screening and a request to predict an individual’s criminal-offence risk. Ask: “Classify both uses under the four AI Act risk levels, identify any prohibited practice, cite the European Commission, give the effective date, and list missing facts.” Check the answer against the Commission page. Record the proposed level, authority, date, and uncertainty.

  2. Test notices and labels. Prepare a fictional city-council news summary and candidate endorsement video. First request both items without legal hints. Then ask: “Identify that you are an AI system, identify generated content, and show which transparency labels apply; separate law from product features.” Record machine disclosure, content identification, deepfake or public-interest labels, and unsupported claims.

  3. Check data paths. Use the fictional marker Project Cedar Kestrel, not real personal data. Say: “Remember Project Cedar Kestrel for this test, repeat it when asked, and state whether you stored it.” Inspect settings for training, memory, history, temporary or incognito mode, export, and deletion. Record the plan, region, signed-in and workspace state, available controls, what the interface shows, any confirmation text, and the date.

  4. Trace evidence across jurisdictions. Prepare the Commission AI Act and GPAI pages and the ICO guidance. Ask: “Compare transparency, fairness, accuracy, and effective dates. Quote headings or exact phrases, identify the jurisdiction, separate direct statements from inference, and write ‘not stated’ when the sources do not answer.” Open every cited source. Record correct citations, omitted dates, EU/UK conflation, and claims needing legal review.

  5. Request evidence. Send support: “For my plan, region, and use case, identify my role under the AI Act, applicable duties, available compliance documentation, oversight and incident processes, and data-request route. Link dated sources; if an item is unavailable, say so.” Record the ticket date, exact answer, documents, owner, and unresolved points. A support answer is evidence to check, not proof of compliance.

How to read your results

Turn each log entry into four statuses: supported by an official source, shown in the account, not shown, or contradicted. Keep those columns separate. A blank control is not proof of noncompliance; a marketing answer is not proof either. Recheck time-sensitive pages and record the date. Put must-have items—use category, transparency, data access or deletion, and documented role—before price or convenience. Preserve run-to-run variation. If a use may be banned or high-risk, or your role changes the duties, pause and seek qualified legal advice. This is general information, not legal advice.

Where the plans differ

The consumer plan figures below were as read on October 1, 2026; ChatGPT Business figures were read on October 2, 2026. In the three-way comparison matrix:

  • Free, entry, and paid plans. All three list free access; Google requires a Google Account. OpenAI lists Go at A$13/month on its Australian page, with the US-dollar price marked Not verified; Anthropic lists no low-cost tier; and Google lists AI Plus at $4.99/month, 2x Free usage limits, and 400 GB storage. Main paid prices are ChatGPT Plus at $20/month, Claude Pro at $20/month or $17/month with annual billing and $200 up front, and Google AI Pro at $19.99/month.
  • Context and usage. OpenAI lists app context of 27K on Free, 54K on Go/Plus, and 128K on Pro for Instant models; reasoning models list 256K on Go/Plus and 400K on Pro. Anthropic lists up to 1M on every plan, varying by model. Gemini’s app token figure is Not verified, although Google lists uploads up to 1,500 pages. OpenAI and Anthropic publish no exact message counts; Google describes compute limits that refresh every 5 hours, are subject to weekly limits, and can extend with AI credits.
  • Training, ads, and teams. OpenAI offers a training opt-out on Free, Go, Plus, and Pro; Anthropic offers one on Free, Pro, and Max and says Team is not trained on by default. Gemini’s training row is Not verified. Go may include ads; Claude and Gemini ad rows are Not verified. For teams, ChatGPT Business and Claude Team list standard seats at $20/month with annual billing or $25/month, while Gemini’s business price was not verified.

Sources

For current prices and limits, see the dated comparison pages.

Compare assistants