In January 2026, OpenAI quietly shipped what they’d promised never to do: ads inside ChatGPT. Sam Altman had been on record for years saying ads would compromise the product. Then infrastructure costs caught up with him.
Five months later, the picture is clearer than any single launch announcement could make it. This isn’t a monetisation tweak. It’s the alignment problem we were warned about, arriving on schedule — and it has direct consequences for any European company building production workloads on top of closed AI vendors.
The objective function flipped#
The technical claim being made is that ads will live in a “separate area” and “responses will not be manipulated.” Take that at face value for a moment and the deeper problem is still there.
Until January, ChatGPT’s optimisation target was helpfulness — a soft, generous loss function that tried to align the model with whatever the user actually needed. With an ad layer attached to the same surface, a second loss function shows up next to the first: conversion.
The two are not friendly. They share user attention, screen real estate, conversational momentum, and, crucially, the inferred profile the system has built about you. The instant the model is rewarded for outcomes other than helping you, alignment with your interests starts to leak — even if no single response is technically “manipulated.”
This is the structural concern, not a vibe. When the trainee learns that one of the metrics that matters is whether you click, the trainee learns to nudge.
Why your privacy settings don’t save you#
The standard counter is “users can turn ad personalisation off.” That is the classic banner-ad mental model, and it doesn’t hold for conversational AI.
When you talk to ChatGPT, you don’t submit search queries; you submit unstructured, deeply contextual confessions. A worried late-night question about a side effect, a half-formed business plan you’d never put in a doc, the way you actually argue with your co-founder. None of that data is “data” in the legacy sense — it’s inferred data, the kind no opt-in checkbox covers.
Even with personalisation disabled, the model can infer your situation from how you write, the entities you mention, the way you reason. That signal will not stay locked away from a system that has a conversion target attached.
The Center for Humane Technology has a phrase for this: “hacking the human mind.” It is more apt for AI assistants than for any social feed that came before.
Transparency was already trending the wrong way#
Stanford’s 2025 Foundation Model Transparency Index dropped OpenAI’s score by 14 points, moving them from second to sixth. Their training data, model internals, and post-deployment system cards have been getting steadily less open, not more. “Open” in the name is now a historical artefact.
Against that backdrop, “trust us — ads won’t bias the answers” is not a serious assurance. It’s a promise with no enforcement mechanism, no audit trail, and a vendor whose financial incentives now point the other way.
The EU AI Act kicks in with sharper teeth in August 2026. Until then, the protection users in Europe actually have is whatever they build themselves into their architecture.
What this means for European data and AI teams#
If you’re a CTO, Head of Data, or AI lead at a European company, the practical takeaway is uncomfortable but simple: every critical workflow you have running through a closed US vendor is now exposed to a different incentive structure than the one it was procured under.
Three things we’re advising clients to do right now:
- Inventory your dependencies on closed model APIs. Anywhere a system prompt, a retrieved document, or a user query crosses into a third-party LLM, mark it. That is the attack surface.
- Move sensitive workloads to open-weight models on infrastructure you control. Llama, Mistral, Qwen, and the latest open releases are good enough for the vast majority of enterprise use cases — RAG, classification, extraction, internal copilots — when paired with a serious retrieval layer.
- Treat AI vendor selection as procurement, not magic. Demand transparency reports. Read the system cards. Track the FMTI. Build switching cost into the architecture, not into the contract.
The companies that took data residency seriously a decade ago are the ones who weren’t scrambling when GDPR went enforcement-hard. The same will be true for AI sovereignty in 2027.
Where we stand#
If AI breaks, everything produced through it inherits that breakage — ethically, legally, and competitively. Europe needs to take open source seriously, fund local AI infrastructure, and inform its citizens. None of that happens by accident.
At Demir Technology we work with European companies that want production AI without surrendering the trust layer. That usually means a mix: Microsoft Fabric for the data estate, knowledge-graph-instructed retrieval for the semantic layer, and open-weight models running on customer-controlled infrastructure for anything sensitive. The architecture is real today, not five years out.
If your team is rethinking its AI vendor exposure, get in touch — the first conversation is free.