Moving from Trust & Safety into AI safety

If you've written policy, run investigations or kept a moderation team calibrated, you already do a version of most AI safety jobs. The open roles show it. Anthropic files its safeguards analysts by harm area, the way an investigations team is organized, and OpenAI's model policy managers do what a policy manager does, for a model.

What you need to add is narrower than it looks, which is good news for anyone who just read a job posting and felt their stomach drop. Here's where each background lands.

What carries over

• Policy writing carries into model policy. Add reading model answers in bulk, and writing rules a model can be trained and tested against.
• Investigations or threat intelligence carry into safeguards enforcement and misuse investigations. Add how API accounts work, and reading a long conversation for intent.
• Quality review and calibration carry into evaluations and red teaming. Add writing test prompts that find failures, and scoring answers the same way every time.
• Operations or program management carry into AI safety program and TPM roles. Add the release process: what gets tested before a launch, and who signs off.
• Engineering or data carries into safety engineering. Add how models are fine-tuned and evaluated.

Close these gaps first

• Use the models every day, and try to break them. Keep notes on what you tried and what got through. That's the start of a red teaming portfolio, and the most fun homework this field will ever give you.
• Learn six words: system prompt, fine-tuning, input and output classifiers, eval (a fixed test set a model is scored against), jailbreak (a prompt that gets around the rules), and agent (a model that can take actions).
• Read the published usage policies of three AI companies side by side, and note where each draws the line. It's the nearest thing to a model policy writing sample you can make before you have the job.
• Take one case from your own work and rewrite it as if the abuse had come through a chatbot. What would you have needed to see, and where could it have been stopped?
• Lead with the harm you know best. Labs hire specialists by harm area: child safety, cyber, fraud, extremism, self-harm. Depth in one is worth more than a general interest in AI.

What it pays

It depends most on who's hiring. On the board's own listings in September 2026, US base pay at P3 was about $293,500 at a frontier lab and $205,720 everywhere else. At P4 it was $348,750 against $260,225.

Google, DeepMind included, pays close to the board's median, so the premium comes from working at a lab. The words "AI safety" in a title don't carry it on their own, so don't rename yourself on LinkedIn and wait for the money.

What one move might look like

Take a fraud investigator moving to a safeguards analyst role. Month by month, it might go like this.

• Month 1: start using two AI assistants daily, and log every refusal and every miss.
• Month 2: read three companies' usage policies and write a page comparing their lines on fraud and scams.
• Month 4: rewrite two of your own fraud cases as if the scam scripts came from a model.
• Month 5: apply for safeguards roles in your harm area, with the comparison and the two cases as your writing sample.

No new degree and no coding course. You keep your specialty, fraud, and show you can do the same job on a new surface. That's the whole trick: keep your harm area, and change the surface.