OpenAI and AI rivals unite on child safety, but will it stick?

A rare moment of industry unity surfaced this week as OpenAI, Amazon, Google, Meta, Microsoft, and five other AI players signed on to Thorn’s Safety by Design principles. The agreement, brokered by the anti-child-abuse nonprofit and ethical tech group All Tech Is Human, aims to weave child safety into every stage of generative AI’s lifecycle — from sourcing training data to deploying models and performing ongoing maintenance. It’s not a handshake deal. The commitments are specific: scrubbing child sexual abuse material (CSAM) from datasets, reporting any confirmed finds to authorities, and building feedback loops that stress-test models before they ever reach a user. OpenAI was quick to point out it already sets age restrictions on ChatGPT and works with the National Center for Missing and Exploited Children. But this pact formalizes what, until now, has been a patchwork of individual company policies.

The challenge, as any veteran of platform safety efforts knows, is that principles are easy to announce and brutal to enforce. Generative AI introduces a terrifying new vector: realistic, AI-generated CSAM (what the group calls AIG-CSAM) created not by scraping the dark web but by prompting a model. The signatories pledge to remove this material from their platforms and invest in research to detect it. Metaphysic, Stability AI, and Civitai joining the effort is particularly notable — open-source image generators have been ground zero for misuse. The working group also committed to annual progress reports, a transparency mechanism that will either build trust or provide a paper trail of broken promises.

Chelsea Carlson, a spokesperson for the effort, put it plainly: ‘We care deeply about the safety and responsible use of our tools, which is why we’ve built strong guardrails and safety measures into ChatGPT and DALL-E.’ That quote covers OpenAI’s portion, but the broader industry statement is notable precisely because it includes competitors who rarely agree on anything — Anthropic and Mistral AI sharing a commitment with Meta and Google. The principles also push for developer accountability, a nod to the fact that API access means the company building the model isn’t the only party that needs to behave responsibly.

Still, I’m watching for two things. First, whether ‘responsibly source our training datasets’ means anything beyond what companies already claim to do, given how opaque data sourcing remains industry-wide. Second, whether the annual progress updates contain real metrics — numbers of reports filed, images removed, models retrained — or just another round of carefully worded blog posts. The technology is moving faster than any working group. The question isn’t whether these companies mean well. It’s whether a set of principles, even specific ones, can keep pace with a problem that scales at machine speed.