Close Menu
  • AI
  • Content Creation
  • Tech
  • Robotics
AI-trends.todayAI-trends.today
  • AI
  • Content Creation
  • Tech
  • Robotics
Trending
  • Tesla Rolls Out FSD V15 with Actuality View and Imaginative and prescient Upgrades – Unite.AI
  • Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI
  • OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI
  • A Group of Researchers from Columbia and CertiK Introduce TensorLean for Formal Verification of GPU Kernels
  • Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI
  • Dragos Finds Immediate Leaks and Stolen Keys in AI Provide Chains
  • Lean and Mathlib Communities Certify Claude-Assisted Riemann Proof – Unite.AI
  • AWS Routes HyperPod Inference Throughout Areas From One Endpoint – Unite.AI
AI-trends.todayAI-trends.today
Home»Robotics»What the EU AI Act Doesn’t Tell Founders About Training Data – Unite.AI

What the EU AI Act Doesn’t Tell Founders About Training Data – Unite.AI

Robotics By Gavin Wallace12/08/20266 Mins Read
Facebook Twitter LinkedIn Email
Gemini Robotics: AI Reasoning Meets the Physical World
Gemini Robotics: AI Reasoning Meets the Physical World
Share
Facebook Twitter LinkedIn Email

Article 50 of the EU AI Act The AI system must disclose to a human that they are talking to an artificial intelligence. The law does not specify what will happen to the conversation after disclosure.

The founders of products built on a model foundation focus on accuracy, and on the price. When selecting vendors, the question of what will happen to users’ data after it has left their product is rarely asked.

The rule became enforceable on August 2, 2026Fines and penalties for non-compliance can reach €15 million or 3% of global annual turnoverThe higher of the two. The scope of this application does not include retention or training. The data practices of the third-party provider will become a part of your software if you send user conversations through your app.

The industry default is to train, not ask

Researchers Stanford’s Institute for Human-Centered AI We reviewed six of the largest US AI companies, including Amazon Anthropic Google Meta Microsoft OpenAI and Microsoft. Each of them uses customer chat data to train models by default, while some retain that data for an indefinite period of time. Jennifer King, the lead researcher at the research institute, was direct when she answered whether AI chat users should be concerned about their privacy. “absolutely yes.” If a document from a service provider does not state otherwise, then training should be assumed. However, the details vary widely from provider to provider and require close scrutiny before making a decision.

Once lauded as a privacy-conscious company,  Anthropic changed its consumer terms So that by default conversations with Claude can be used to train users, but they have the option of opting out. The training policy of a vendor can be changed very quickly, even though privacy used to be part of the original pitch.

Other models don’t allow you to opt out. Google keeps Gemini conversations reviewed by human annotators for up to three yearsGemini Apps Activity is not able to undo the effect. Google keeps a copy of a conversation in its systems for a certain period after it has been flagged as requiring human review. This is regardless of how the user uses that conversation.

Character.AI trained on user conversations by default The company only offered opt-outs to UK and EEA users, putting it on the more permissive side of data practices. The company also faced reputational and legal challenges. The company will be in January 2026. Character.AI and Google agreed in principle to settle a group of lawsuits The platform was accused of harming minors through its chatbot interaction with the platform, which is unrelated to data training policy. Google, which has ties with Character.AI, was also named as co-defendant. While no liability or terms were disclosed, this case demonstrated that the legal risk associated with an AI vendor could reach companies who back it.

Subcontractor access, retention periods, and human review triggers are rarely mentioned outside of the fine print in a policy. Integrating these APIs requires companies to conduct extensive investigations to discover how long data remains in the review process or who has access to it. Most of this information never reaches the end user, so a company could inherit data practices from a vendor and expose its users to them without realizing.

Why this Is Not Just A Policy Debate, But a Problem With The Founder

Even if it wasn’t your intention, if your product routed user conversations to a foundation model, with default training turned on, you have accepted a data privacy policy without informing your users. This decision is not required to be disclosed by Article 50 because the law only applies to AI.

This exposure is increased by third-party access. Subcontractors are regularly used by model providers to perform data annotations and safety reviews. By May 2026 Texas Attorney General Ken Paxton opened an investigation into Meta’s AI glasses Data annotators working for a Kenyan-based subcontractor, Sama, reportedly claimed that they were able to access private footage from users. This included bathroom visits. Meta denied that the report reflected its entire practices. The investigation continues. The same problem can be applied to conversational AI, even though the example is about wearable cameras. If a business promises that they won’t share your information, it doesn’t mean anyone else will be able to read it.

What you should check before choosing a vendor

Every founder needs to answer a few questions before they can connect a model of underlying foundations with actual user data.

Start with the data processing contract. The data processing agreement is the key document that governs your company’s handling of user information.

Contracts themselves deserve equal attention. The vendor may offer more restrictive terms for no training on API or enterprise contracts than consumer chat, so it is important to confirm the terms that apply. Most companies assume that enterprise-level privacy terms cover them automatically.

Also, opt-out options are worth looking at. Google Gemini’s settings show why this is the case, because in that model disabling activities tracking does not prevent human review. Vendors should clearly state whether opting out stops training or human review.

Lastly, determine where data is sent downstream from the vendor. This is because annotation and safety reviews are often outsourced. A vendor’s privacy promise does not automatically apply to all subcontractors in the chain.

All of these questions shouldn’t be treated like a single-time activity. Vendor conditions can change quickly and you may find that the terms of the vendor with whom you worked last year are not the same today.

The market can set a standard that the regulation hasn’t reached yet

It’s unlikely that Article 50, the EU’s last statement on AI governance will be the end of the road. Founders don’t need to wait until the next set of regulations to fill in the gaps left by this article. The gap can be closed now by implementing a clear and transparent opt-in to train on the data of their users. It’s possible to improve a model without compromising user trust. There are many AI services that still use conversations for training, without ever requesting explicit informed consent. If founders draw the line themselves, they won’t have to worry about retrofitting compliance in the future.

This is a precedent-setting move. The encryption of messaging apps and the password manager didn’t have to wait until regulators made strong defaults a standard. The companies made these choices independently, and they turned it into a strong market position. AI products also have this opportunity. The people who are able to draw the line in advance of it being mandatory will have the trust and confidence from their users for the next conversation that is sensitive.

AI dat data founders training
Share. Facebook Twitter LinkedIn Email
Avatar
Gavin Wallace

Related Posts

Tesla Rolls Out FSD V15 with Actuality View and Imaginative and prescient Upgrades – Unite.AI

22/11/2026

Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI

22/11/2026

OpenRouter Cuts Three Suppliers After AI Immediate Publicity Findings – Unite.AI

22/11/2026

Collectively AI Discloses Immediate Leak After Coding Agent Uncovered Credential – Unite.AI

22/11/2026
Top News

Google DeepMind Staffers Ask Leaders to Keep Them ‘Physically Safe’ From ICE

What is Google One? The Google One Plans and Pricing

Riley Walz joins OpenAI, the Silicon Valley jester.

The DOGE used AI to help with housing policy The government won’t say how

White House Trying to Decide What to do About Chinese AI

Load More
AI-Trends.Today

Your daily source of AI news and trends. Stay up to date with everything AI and automation!

X (Twitter) Instagram
Top Insights

A Coding Solution to Create and Train Advanced Architectures Using JAX and Flax.

11/11/2025

AI’s impact on politics, technology, the media and more

15/09/2025
Latest News

Tesla Rolls Out FSD V15 with Actuality View and Imaginative and prescient Upgrades – Unite.AI

22/11/2026

Potential Cognition AI Spherical Would Elevate $2B-Plus at $75B Valuation – Unite.AI

22/11/2026
X (Twitter) Instagram
  • Privacy Policy
  • Contact Us
  • Terms and Conditions
© 2026 AI-Trends.Today

Type above and press Enter to search. Press Esc to cancel.