DEV Community

The TrustHawk team
The TrustHawk team

Posted on

Before you adopt an AI tool, read its regulator record

Disclosure: this article is by The TrustHawk team, and we built TrustHawk, the site the records below come from.

Most questions developers ask before wiring an AI tool into a product are about capability: context length, latency, price. Fewer ask what the company behind the tool has already been fined for, investigated over or caught exposing. Those records are public, and they say a lot about how a vendor treats data once it has it.

This post walks through what those records look like for some well-known AI companies, using what is on file in public sources, and then turns that into a short checklist you can run before you adopt a tool.

What a regulator record actually contains

There are a few kinds of public record worth reading, and they mean different things.

Privacy fines. In Europe, data protection authorities publish GDPR decisions, and trackers such as the CMS GDPR Enforcement Tracker collect them. A fine is a finished decision, not an allegation.

Regulator actions and investigations. An order to stop processing data, or an open inquiry, is a signal that a regulator has concerns, but it is not a final finding.

Settlements. In the US, Justice Department press releases describe civil penalties and settlements, often over children's privacy.

Breaches and exposures. Breach databases such as Have I Been Pwned, and incident trackers such as the AIAAIC Repository and the AI Incident Database, record exposed data and reported harms. Incident database rows are reports, not proven findings.

Some records on file

OpenAI. The CMS GDPR Enforcement Tracker lists a fine from the Italian Data Protection Authority (Garante) dated 2 Nov 2024, for non-compliance with general data processing principles. On the policy side, OpenAI's privacy policy says it may use the content you provide "to train the models that power ChatGPT", and it publishes instructions for opting out. So training is on unless you turn it off. OpenAI also states it holds a SOC 2 Type 2 report and ISO/IEC 27001:2022 certification for its API and business services, and it runs a bug bounty through Bugcrowd.

DeepSeek. The AIAAIC Repository records a report from 29 Jan 2025 that a Wiz researcher found a publicly open DeepSeek database, listed as exposing user data and chat histories. The next day, 30 Jan 2025, Italy's Garante ordered DeepSeek to limit processing of Italian users' data and opened an investigation. On 17 Feb 2025, Korea's Personal Information Protection Commission examined DeepSeek, and app downloads were suspended in Korea on its recommendation. The OpenAI vs DeepSeek comparison puts the two companies side by side, and the full source list is on the DeepSeek company page.

Character.AI. The Garante fined Character.AI on 3 Jul 2026 for non-compliance with general data processing principles. Separately, on 18 Aug 2025 the Texas Attorney General opened an investigation, by civil investigative demand, into Character.AI over deceptive mental health claims. The same Texas investigation also named Meta AI Studio. The Character.AI company page links each of these to its source.

Replika. The Garante fined Replika on 10 Apr 2025 for non-compliance with general data processing principles. Its privacy policy does not say that training on your data is opt-in only, and we could not confirm a way to opt out of training from that policy.

Microsoft and Amazon. Both have US children's privacy settlements on record. On 12 Jun 2023, Microsoft agreed to pay a $20 million civil penalty for alleged violations of children's privacy laws. On 19 Jul 2023, Amazon agreed to injunctive relief and a $25 million civil penalty for alleged violations of children's privacy law relating to Alexa.

Meta and Suno. Have I Been Pwned lists an Instagram breach dated 7 Jan 2026, and a Suno breach dated 25 Nov 2025. Meta also has Irish Data Protection Commission fines on record, including one dated 17 Dec 2024 for insufficient technical and organisational measures to ensure information security.

What these records mean if you are choosing a tool

A few patterns come out of reading records like these.

Italian regulators move early on AI products. The Garante shows up again and again: OpenAI, DeepSeek, Character.AI and Replika all have Garante decisions or orders on file. If you serve users in the EU, a Garante decision is a reasonable early warning that a product's data handling may not meet GDPR expectations.

Exposure and enforcement often arrive together. In DeepSeek's case, a reported open database was followed within weeks by action from regulators in two countries. A breach report on a vendor you depend on can quickly become a compliance problem for you, not just for them.

Children's privacy is where US penalties land. Several of the US settlements above are about children's data. If your product could be used by minors, check the vendor's record here specifically.

An empty record is not a clean record. A company with nothing on file may simply not be covered by the sources, or may be too new to have a history. Treat "no events found" as "no evidence either way", not as a pass.

A short checklist before you adopt an AI tool

  1. Search the GDPR enforcement trackers for the company's legal entity names, not just the product name. Fines are often filed under an Irish or other EU subsidiary.
  2. Read the training clause in the privacy policy. Is your data used for training only if you opt in, or is it on by default with an opt-out? For business or API plans, check whether the terms differ from the consumer app.
  3. Check for export and deletion. Can users get a copy of their data and delete it? You will need this for your own data subject requests.
  4. Look for a security programme you can verify. A bug bounty or security contact, and a SOC 2 Type II report or ISO 27001 certificate that covers the product you will actually use, not a sister product.
  5. Search breach and incident databases such as Have I Been Pwned, the AIAAIC Repository and the AI Incident Database. Remember that incident rows are reports, not findings.
  6. Check open investigations, not just fines. An open inquiry today can become a restriction on the product in your users' region tomorrow.
  7. Write down what you found and the date. Records change, and a vendor review from last year may already be out of date.

None of this decides for you. It shows what has already happened, in public, with data people trusted the company with, which is worth knowing before you hand over your users' data too.

The TrustHawk team

Top comments (0)