Ethics, built in, not bolted on

"It's just a tool" might be the most load-bearing sentence in technology, and I've come to treat it as a warning sign. Not because it's entirely wrong; a model doesn't have intentions, but because of what the sentence is for. Nobody says "it's just a tool" while designing a system. They say it afterward, when a consequence has arrived and no one wants to own it.

Here's the problem: a deployed AI system is not a hammer. It's an accumulation of decisions: whose data trained it, what objective it optimizes, who it was tested on, where it got pointed, what happens when it's wrong. Every one of those decisions had an author. "It's just a tool" is a machine for making the authors disappear.

Our sixth commitment exists to keep the authors in the room.

Harm arrives through defaults, not villains

The best-documented AI failures share an uncomfortable feature: nobody involved was trying to cause harm. Amazon built an experimental resume-screening model and discovered it had taught itself to penalize resumes containing the word "women's," as in "women's chess club captain," because it was trained on a decade of hiring outcomes from a male-dominated industry. To Amazon's credit, they caught it and killed the project. But note what happened: the system hadn't malfunctioned. It had faithfully learned exactly what the data taught it.

ProPublica's 2016 analysis of COMPAS, a recidivism-scoring tool used in sentencing decisions, found that Black defendants who did not go on to reoffend were nearly twice as likely as white defendants to have been misclassified as high risk. The vendor disputed the framing, and the academic fight that followed established something genuinely important: several perfectly reasonable definitions of "fair" are mathematically impossible to satisfy at the same time. Which means fairness in these systems is not a property you verify. It's a trade-off you choose, and if you didn't choose it consciously, you still chose it.

That's the pattern worth internalizing: in AI systems, harm doesn't usually arrive through malice. It arrives through defaults: inherited data, unexamined objectives, deployment contexts nobody analyzed. Good intentions are not a control. Reviews are controls.

Built in, not bolted on

Which is why the mechanics of this commitment are deliberately unromantic:

• Every AI use case gets a risk, IP, and bias review before it rolls out. Before, not after. An ethics review conducted downstream of the build is a compliance exhibit, not an engineering input. Frameworks like the NIST AI Risk Management Framework make useful scaffolding, but a framework you consult at the end changes nothing about what you built. The review has to sit where the design decisions get made.
• We say the trade-offs out loud. There is no free lunch in this field, and we won't pretend otherwise to close a deal. More automation brings skill-atrophy and oversight risks. More model capability means more data-exposure surface. A vendor who presents an AI rollout with no cost column is either not thinking or not telling you.
• An engineer can raise a harm concern without betting their career on it. The engineering profession's grimmest case studies, including the engineers who warned about Challenger's O-ring seals the night before launch and were overruled, are mostly stories about organizations that made honesty expensive. The person closest to the system sees the failure mode first. If raising it is dangerous, your ethics process is decorative, whatever the policy document says.
• We walk away from work that would harm users, profitable or not. This one has no clever mechanism behind it. It's a decision made in advance, so it never has to be re-litigated under commercial pressure, which is precisely the condition under which it would lose.

Owning our share

The commitment says we own our share of the impact while we create real value, and both halves matter. We are not neutral about AI; we build with it every day, and I believe the value is real and large. That's exactly why the honesty matters. The people who will poison this well aren't the skeptics; they're the boosters who ship harm behind "it's just a tool" and let the trust drain out of the entire field. Naming trade-offs, reviewing before deploying, and facing bias directly isn't hedging our enthusiasm. It's what taking the technology seriously looks like.

Sources

• Jeffrey Dastin, "Amazon scraps secret AI recruiting tool that showed bias against women", Reuters, October 2018.
• Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner, "Machine Bias", ProPublica, May 2016.
• NIST, AI Risk Management Framework, 2023.