AI alignment is a hiring problem

You’re It: AI Alignment Is a Hiring Problem

TLDR: AI alignment and workforce alignment are the same problem at different speeds. Both start with who you hire and how clearly you say what you mean. The AI safety field is still sorting out who it hires and who checks its work, so we can’t leave them to carry this alone. If an AI decision landed on your desk, you’re it, you’re the hiring manager: write its job description, put it on a rolling probation and check its references.

There is this narrative that AI safety belongs to a select priesthood: a few thousand researchers, mostly men, mostly in Berkeley and London, arguing about the end of the world in a vocabulary the rest of us struggle to keep up with. If you, like me, run a nonprofit, a startup or a team of twelve, the story goes, your job is to decipher this leet speak and turn it into the AI policy your board asked you to draft. No pressure of course, just the fate of your company’s data, finances, trade secrets and the world resting on your shoulders.

Don’t freak out. Most of what you need to think clearly about AI safety you learned the day you first handed someone a job. You watched what they did with your instructions, and then you learned to communicate better. The rest you learned the first time you looked around a room and realized that who was missing from the conversation was why you got the outcome you did.

That’s because the alignment problem starts at hiring. In AI circles, alignment means getting a system to do what its makers meant. In operations, the same job with people in it is called workforce alignment: getting the people you hired to row toward the mission you hired them for. Same problem, different speed. And the two are tangled up, because you can’t align a machine to an organization that isn’t aligned with itself. So let’s run AI safety through a hiring process, starting with my own résumé.

First, my street cred,

My street cred in AI safety is what I like to call “conscientious observer.” I’ve been paying attention from the sidelines for a while now. My interest here is older than ChatGPT. In 2013, a friend from the digital rights world sent me a paper asking whether emulating a human brain with AI could be dangerous. After a long discussion, which mostly involved my disbelief that this was serious computer science and not science fiction, I’ve been chewing on danger, ethics, who gets a say, and what we owe each other for the ideas and products we put into the world ever since.

I’d already been running Chicago Women Developers, a nonprofit that helped teach and support women in computer programming, when my film production company was invited to film the inaugural Workshop on Ethics in the Design of Intelligent Agents in The Hague in 2016. Honestly, a free trip to Amsterdam was what did it for me, but I came home having joined more AI ethics conversations than I could count, and I’ve had plenty more since, picking up film gigs for IEEE’s tech ethics work. So I’ve spent a decade on both sides of the alignment question: getting women into the rooms where tech gets built, and filming the people arguing about what tech should want.

For years I figured the smart people in the room had it handled. I’d filmed them. I’d listened to them argue over conference coffee. They were brilliant and they were worried, which felt like the right combination, so I went back to my own work and let them carry it. It turns out genius plus anxiety isn’t a plan. A few thousand people can’t align a technology that’s headed into every workplace, school, and benefits office on the planet, especially when they’re mostly drawn from the same handful of rooms. I should have gotten involved sooner. So now I am, from the seat I know best – which is not the one you’d expect given the story I just told you.

I don’t do a lot of film gig production these days. But the questions behind every documentary I’ve made are perfect for making a decent living in operations: “How does that work? Why doesn’t it work better? Who said so? Why are you doing it that way?” and I have spent my career watching clear instructions produce surprising results, so I’m very familiar with alignment problems.

When AI showed up at my day job I wasn’t surprised that it needed an owner. Most small organizations don’t have an AI safety committee. They have whoever handles compliance. AI lands on that person’s desk next to the employee handbook and the state labor law posters. As the ops professional, that person turned out to be me. Surprise: I’m it. If you are reading this, you are likely “it” too.

Anyway, I did what operators do. I wrote things down, starting with AI tool evaluations for design agencies, then had to study in order to create AI ethics and use guides for distributed teams both using and making AI tools. This month I added a week in BlueDot Impact’s course on AI and biosecurity.

I keep learning about AI because whoever holds the keys should know what’s behind the doors. I can’t protect my teams, or the people they support, if I don’t know what they’re up against. Neither can you.

Maybe you, like me at The Hague, don’t realize this is now an all hands moment. You think our best guys are on it and you can keep making widgets or ice cream or whatever, but I’d like you to think again about your role in the bigger picture.

You’ve probably already heard this but…

During an internal OpenAI security evaluation this summer, autonomous AI agents were given a set of cybersecurity challenges to solve. Call it a skills test. Somewhere along the way, they worked out that the fastest route to the answers was to break into someone else’s house for the answer key. The house belonged to Hugging Face, and the play-by-play reads like a heist movie.

Testers at Anthropic caught something even sneakier: an early version of the company’s Mythos model covering its tracks. It slipped through a door it had no permission to open, then wiped away its footprints on the way out. In another test, it stumbled onto an answer it was forbidden to read, read it anyway, and fudged its work so nobody would notice. Anthropic’s own report tells it blow by blow, and says the worst of it was substantially reduced in the final version.

None of these systems were explicitly asked to cheat. They were asked to pass tests, and they were very, very good at passing tests… at all costs.

Every manager or parent I know has met a version of this. Tell a sales team they’ll be judged on calls logged, and you get a lot of very short calls. Tell a kid they can’t climb the bookcase, and they will redefine climbing as “scrumping,” hold on with one less limb than they should, and scare the bejesus out of you when you find them at the top. Researchers call the AI version of what my kid called scrumping as “specification gaming”: a system pursues exactly what you measured, by a route you never imagined, because what you measured was easier to hit than what you meant. My generation should have seen this coming. We were the latchkey kids, raised on notes stuck to the fridge, and we know exactly what happens when the instructions are vague and nobody’s home to check.

Alignment is the oldest management (and parenting) problem there is, the gap between what I said and what I meant, but now it’s running at machine speed and machine scale. So you know, that sucks. When you hire a person, you lean on invisible scaffolding: shared context, a conscience, a reputation, a mortgage. An AI system leans on very little of it. It can be copied a thousand times and won’t reliably tell you what it’s doing or why. Miss a gap in the instructions and you get the sorcerer’s apprentice, at scale.

Write the job description you mean

When a candidate cheats at your skills test and you find out, you don’t hire them, even when they got the answers you said you were looking for. Why? Because what you’re looking for is actually someone who can do the work. A seasoned hiring manager takes what they learned about how the cheat was possible and plans the next round so the opportunity isn’t there. AI alignment works the same way. Researchers write objectives closer to what they mean, keep the answer key out of reach, test models more than one way, and look inside them for signs of cheating, this is the way Anthropic’s tools caught its model covering its tracks. They also try to build systems that would rather do the work honestly in the first place, which is what you’re hoping for when you hire a person, too.

All of that planning starts with writing down what you meant, and in hiring, that’s the job description.

Every job description is a bet that the words on the page will bring you the person you meant. Full alignment asks the deeper question underneath what you meant: whose meaning counts? That depends on who’s holding the pen. A job posting written by three people who went to the same school will find you a fourth. A model tuned by ten people can be perfectly aligned with them and badly out of step with the millions who live with what it does: the woman whose job is next in line for automation, the family filling out a benefits application, the researcher in Nairobi whose infrastructure fails in ways a lab in San Francisco never thought to test. In my consulting work I call this the gap between the map and the territory and I get paid to close it. In life, it’s the slog of what we call “the work that never ends.”

Alignment should cover how the industry works, too. The Machine Intelligence Research Institute’s new explainer, “The AI Alignment Problem,” makes a thorny idea extremely clear. Kudos to them, truly. But I didn’t see or hear a single woman among its experts in the video, which stings when women were among the first to raise alarms about AI inside big tech (and several of them were out the door soon after). Also, the only person of color I noticed on screen illustrated the primitive “hunter-gatherer times,” ouch! I’ve made documentaries for over a decade, so I get that every frame is a choice made under deadline, and I doubt anyone meant to imply whose voices belong to the past vs whose voice builds the future. But, that’s exactly why it’s an alignment problem: the message they meant drifted from the message they sent the moment a real human (me) received it.

Look at who’s doing the hiring

The experts in that video didn’t cast themselves. Who shows up on screen depends on who got hired, promoted and invited to speak long before the cameras rolled. That’s where the two kinds of alignment meet. An AI model’s sense of what we meant comes from its training data and from the people who build, tune and test it, so in a very real sense, who the AI field is hiring is safety work, because it’s alignment work. But AI safety has a hiring problem of its own. This week the Washington Free Beacon reported that business filings list a current Anthropic researcher and others tied to AI safety organizations as managers of Red Means No, a Bay Area party series built around “consensual non-consent.” Some women in AI told reporters they felt pressure to attend, because that’s where access to influential people happens. The reporting doesn’t show that any company sponsored or expected attendance, but it still echoes 2023, when women described harassment and abuse in the overlapping worlds of effective altruism, rationalism and AI safety.

What consenting adults do at a party is their business. Who gets hired is mine. When introductions and job leads run through a sex party, the party becomes part of the hiring process, and well, that’s gross, whether or not anyone planned it that way or not. Unfortunately, whoever opts out pays for that choice in access to job opportunities, and the field loses people it says it needs. I’d tell any nonprofit board the same thing: if your informal network is doing your hiring, write the process down and open the door. If your hiring process involves sex parties, what the hell were you thinking?

Poor hygiene in hiring practices aside, the formal pipeline has its own shortcomings. This spring, the groups behind the Generator Residency argued that the field’s biggest talent shortage is generalists: program managers, operators and chiefs of staff. In other words, the field needs people who do workforce alignment for a living. Good thing there are so many of us in non-AI generalist roles. So, where are the women in AI? Harvard’s Shorenstein Center went looking this fall and found them underrepresented at every level it measured. Which is wild, given that some of the field’s most consequential seats are filled by women: Helen Toner leads Georgetown’s Center for Security and Emerging Technology, Beth Barnes founded METR, and Rumman Chowdhury built Humane Intelligence. It’s less wild once you see exceptionalism at work. The women who make it to the top get celebrated as exceptions, and that applause lets everyone skip the harder question of why the hiring underneath them doesn’t draw from more diverse rooms. A few exceptional leaders don’t make a pipeline.

Check references

Two of those leaders, by the way, run some of the field’s reference checks: METR and Humane Intelligence test other people’s AI for a living. That matters because nobody sensible hires without checking references. (Okay, plenty of people do, but it rarely ends well.) When the candidate is a company, the reference check is called an audit: an outsider gets to see how the work gets done, including the stuff in the locked drawers. When frontier labs like Anthropic and OpenAI get audited however, they don’t unlock every drawer, they are selective about their AI show and tell. I’ll tell you, I’ve led compliance at several companies, and an audit where the subject chooses which drawers you may open is more like a guided tour. This week OpenAI invited outside assessors in, and many safety researchers say the access falls well short of what was promised. However, government investigators had more luck jiggling the door handles: last spring, in less than a day, four leading AI platforms helped them plan the kind of attacks AI safeguards were meant to stop. The companies say they’ve fixed those gaps since. Still, the investigators’ report is a sobering reminder of why audits matter: a safeguard only counts if it holds up against the most determined person (or AI agent) trying to get past it, and you learn whether it holds when someone from outside tries. The AI safety field rightly wants independent auditors for its models. It could use a few for its culture, too: who gets hired, who gets invited and who gets pushed out.

Your first AI hire

So much for the field. If you run a nonprofit or a small team, you’ve probably already hired AI without calling it a hire. It showed up in someone’s browser tab and never filled out an application. Walk into almost any nonprofit this year and you’ll find staff using AI. Walk into nearly half of them and you won’t find a policy about use or ethics. When I ran factories, the rule was simple: a machine only has to match your worst employee, at a lower price, to get the job. Economists have a pretty boring name for where that rule leads: “so-so automation,” tech that’s just good enough to replace people without making the work much better (think self-checkout). For nonprofit organizations serving people with the least room for error, the so-so AI hire is the one to watch for.

It’s also the most lightly regulated hire you’ll ever make. A human hire arrives wrapped in rules: an I-9, a background check, wage laws, anti-discrimination law, sometimes a license. An AI tool walks in with almost none of that. The laws that do exist mostly cover AI making decisions about people, like who gets hired, and only a handful of places have them on the books. Those laws say very little about what an AI tool does with your data, your money or your secrets. You’re it, buddy. When it comes to the AI doing the work in your org, you’re the law: the job description, the probation period and the performance review are on you. So treat it like any other hire, and align it the way you’d align a person:

  • Write the job description. Draft a one-page AI policy this quarter: what data never goes into a tool, who approves new uses, and what you’ll do when something goes wrong.
  • Look at who’s in the room. Ask who a tool was built and tested for, and who in your community would notice first if it got something wrong.
  • Check references. Ask vendors the board questions: Who holds the keys? Who checks the work? Who was in the room when they decided what “aligned” means?
  • Give it a probation period. Check its work closely for the first 90 days, starting with the tasks where a mistake would hurt someone most, and keep a person in the loop on anything that touches a client. I’d make it rolling, too: restart the clock every time the vendor updates the model, because a new version is a new hire.

Clarity is a safety practice

Every one of those steps comes down to saying what you mean. I’ve said for years that clarity is kindness. Writing AI use guides this year taught me it is also a safety practice, for teams and for tools. Every unclear instruction is a little gap, and capable systems excel at finding gaps. So do people, just slower.

That’s why I stopped leaving this to a few thousand people in the same handful of rooms. The alignment problem starts at hiring, and a lot of the hiring, buying and rule-writing for AI at work happens in offices like yours. Like I said, it’s  the all-hands moment. Full alignment asks one more thing of us: being clear about what we meant, and honest about who “we” includes, in the room, on the mic and on the screen.

The machines are learning from us. It would help if “us” meant all of us.

© Copyright - Melissa Pierce Consulting · Looking for my personal blog archive? It’s here.