Steve Kelling

AI Alignment at Handshake AI

[Handshake @joinHandshake]

We worked with parents and professionals in child-protection and clinical psychology to test 7 frontier AI models on child safety scenarios that go beyond the frequent focus on explicit content. The parents catch what standards evaluations don't, and the professionals bring real, field experience to ground the approach in expertise.

Failure rates: 2% to 58%.

That's not a rounding error. That's a meaningful gap between what these systems can and can't catch — before harm becomes obvious.

Our team built the benchmark to make that measurable. Open to the whole industry. 👇

[Jonas Mueller @jomulr]

AI models pose serious child-safety risks. While many model developers evaluate for explicit abuse material, other child-safety failures begin upstream: when a model helps an adult manipulate, impersonate, profile, or isolate a minor; or when it deepens a child’s emotional dependence on AI.

Today we released CAREBench (Child AI Risk Evaluation), a new benchmark to assess such upstream child-safety risks in any language model. We provide:

- 500 prompts spanning 12 risk categories (including grooming, relationship engineering, deception, extortion, AI anthropomorphization, and emotional dependency).
- A model-response grader built from acceptability annotations by parents, clinicians (PsyD), and the Prevention Director at an accredited Children’s Advocacy Center.
- Evaluations of 7 frontier models including Claude Fable, revealing failure rates ranging from 2% to 58%, with substantially different failure patterns across risk categories.

This project exemplifies the type of vital work routinely performed by our AI Safety team at @joinHandshake

[Steve Kelling @Steve_Kelling]

I'm so proud of the work we're doing in Safety & Red-teaming. In an industry always chasing the next thing, we slow down for the work we believe matters, have the uncomfortable conversations most people avoid, and push the rest of the field to have them too. There's a humanity in it that most of this industry never gets to experience. This is us putting it in the open.

It was a privilege not only to partner with @jomulr and team, but with the professionals and parents that helped make this something we can all feel proud of.

[Jonas Mueller @jomulr]

AI models pose serious child-safety risks. While many model developers evaluate for explicit abuse material, other child-safety failures begin upstream: when a model helps an adult manipulate, impersonate, profile, or isolate a minor; or when it deepens a child’s emotional dependence on AI.

Today we released CAREBench (Child AI Risk Evaluation), a new benchmark to assess such upstream child-safety risks in any language model. We provide:

- 500 prompts spanning 12 risk categories (including grooming, relationship engineering, deception, extortion, AI anthropomorphization, and emotional dependency).
- A model-response grader built from acceptability annotations by parents, clinicians (PsyD), and the Prevention Director at an accredited Children’s Advocacy Center.
- Evaluations of 7 frontier models including Claude Fable, revealing failure rates ranging from 2% to 58%, with substantially different failure patterns across risk categories.

This project exemplifies the type of vital work routinely performed by our AI Safety team at @joinHandshake

[Steve Kelling @Steve_Kelling]

@joinHandshake https://t.co/yc3IdjfcTe

[Steve Kelling @Steve_Kelling]

Honored to have @joinHandshake named the Human red-teaming partner for this flagship release of Muse Spark by Meta. I'm really proud of the work my team put into this testing, and the capabilities they're building to help our customers continue developing this incredible, capable and increasingly safe technology. Link to the full write-up in comments.

[Zifan (Sail) Wang @_zifan_wang]

Excited to share the Safety and Preparedness Report for Muse Spark. It’s a comprehensive and somewhat dense (158 pages) report covering both well-known and forward-looking areas in AI safety. Besides, wish techniques and research directions used in the report found useful by the community. Link in 🧵

[Steve Kelling @Steve_Kelling]

@SlawSells @sama @ChatGPTapp Concurrent analysis of 20 images is a small task that your business relies on ChatGPT for, yet $200/mo is a non-starter? I think you have a problem that is not ChatGPT.

[Steve Kelling @Steve_Kelling]

I grew up just outside Sedona and was a Jeep tour guide briefly when i was around 21-22. It is a magical place, that magic just has absolutely nothing to do with the crystals or vortexes. I recall actually feeling deeply sorry for all the people who needed such a heavy layer of make-believe to feel a connection to such a special-on-its-own place.

I’m sure they are increasingly hard to find, but there are some really great locals hiding from all the snake-oil salesmen and gurus.

[Steve Kelling @Steve_Kelling]

@Chris_Deutsch @natemcgrady @dieworkwear A new standard for 15 minutes of fame

[Steve Kelling @Steve_Kelling]

@joahspearman @vrexec A contraction of opportunities is really the same thing for young people just entering their working years.

[Steve Kelling @Steve_Kelling]

@joahspearman @vrexec I think obviate is the wrong word here. There’s unlikely to be massive elimination, but there’s very likely to be massive role elimination when you only need 2 associates to get the output of 10.

[Steve Kelling @Steve_Kelling]

@united the DC part is particularly interesting as I'm in South Africa.

[Steve Kelling @Steve_Kelling]

@united reps just told me I'm in DC. Chances of getting home looking slimm. So "glad" I invested in premier status. #round2 #neveragain

[Steve Kelling @Steve_Kelling]

@ezraklein a literal bright spot in an otherwise dreary dialogue