How to answer ethical concerns about AI

Adam Faik

Jul 26, 2026

Every time I talk about AI with someone, the conversation ends up in the same place. Not on models. Not on prompts. On ethics. Someone says the data centers are draining water tables. Someone says the training data was scraped from people who never agreed. Someone says they became a designer to design, not to review a machine’s output. The objections that stick aren’t about capability anymore, they’re about conscience.

In early 2025, I watched a French designers’ collective called Designers Éthiques lay out every one of these objections in one 40-minute talk. It’s stayed with me since. Not because it was contrarian, but because it wasn’t. These were rigorous professionals making the strongest version of the case: an eco-design specialist, a design researcher, an ergonomist. No trolling, no panic. It’s the sharpest compact map I’ve found of what your own team members are thinking and mostly not saying in the meeting.

If you’re leading a team through an AI transition, you’ve met these objections too. Maybe in a retro, maybe in a 1:1, maybe as a silence that never turns into usage. And you’ve probably been handed exactly one playbook for them: overcome the resistance. I think that playbook is wrong. Your skeptics are mostly raising real problems, and the fastest way to lose them is to debate.

The better move: sort each concern into one of four kinds. Some are true, and you change what you adopt. Some are risks you control, and you change how you adopt. A few are misconceptions, and you correct them with evidence instead of marketing. One or two are values, and you respect them. The payoff isn’t a converted team. It’s AI use your team can defend out loud.

The playbook, at a glance:

The trap. I’ll show you why winning the argument against a skeptic loses the team.

The sort. Here’s the four-kind triage that replaces the debate: true, fixable, misread, or a value.

The map. You get the seven concerns you’ll hear most, each with an honest verdict, the move that follows, and the one resource worth forwarding.

The practice. What a defensible team AI posture looks like once the sorting is done.

By the end, you’ll have a verdict and a concrete next move for the seven objections coming your way this quarter, plus the one sentence that keeps a skeptic on your team. You stop dreading the ethics conversation and start using it to make your team’s AI practice sharper. No slides required, no philosophy degree either.

Let’s sort this out.

Why winning the argument loses the team

The talk opens with a moment I can’t stop thinking about. At a green-IT conference workshop, in front of the most skeptical crowd available, the speakers ran a session asking “AI: in or out?” Almost every group landed on “in,” reasoning that it’s here anyway, there’s no choice, so let’s make it as clean as possible. One speaker was troubled by exactly that phrasing. ”We have no choice” is not what agreement sounds like, it’s what resignation sounds like. If your team adopts AI in that spirit, you didn’t win them. They just stopped telling you things.

There’s a second reason the debate is rigged before you open your mouth. Your team members already met AI adoption as users, and it wasn’t polite. The design researchers at Limites Numériques documented the pattern: AI buttons pushed front and center, features switched on by default like Strava’s Athlete Intelligence, dialogs that offer “try it” and “not now” but never “no.” Sparkles and purple everywhere, the visual vocabulary of magic. When you pitch AI to your team with vendor enthusiasm, you pattern-match to the forced adoption they already resent. They’ve heard “this will make everything better” before, from a button they couldn’t refuse.

Organizational research has said this plainly for years. In their 2008 Academy of Management Review paper, Ford, Ford and D’Amelio argued that resistance to change isn’t a defect in the resisters. Change agents cause a good share of it themselves, and resistance is better treated as a resource: engagement, feedback, proof that people take the change seriously. The person pushing back is often the person paying the most attention.

If you read How to lead a tech team through the AI shift, you might spot a tension here. I argued there for the 20-60-20 rule: pour your energy into your champions and the watching middle, and stop exhausting yourself arguing with the resistors. I stand by every word, and this article doesn’t change the math. Sorting is not arguing, and what follows is not a conversion campaign. It’s for the conversations that find you anyway: the 1:1 where a concern lands on the table, the team meeting where a hand goes up. You still don’t chase your skeptics; you answer well when they’re in front of you, because the watchers are scoring how you do it. One honest answer to a skeptic moves ten watchers.

And there’s simple arithmetic about credibility. Argue once against something that turns out to be true, say the data-center water numbers, and you lose the room for a quarter. The skeptic came with figures. You came with talking points. So the job isn’t to win. It’s to sort.

Sort the concern before you answer it

Here’s the move that changes the conversation: treat each concern as a claim to classify, not an attack to parry. A triage nurse doesn’t argue with symptoms. She figures out what kind of problem she’s looking at, because the kind determines the response. Four kinds cover almost everything your team will raise.

Kind one: they’re right. The data-center buildout really is doubling electricity demand. When a concern is true, the honest answer starts with “you’re right,” and the follow-up is a change in what you adopt. Nothing builds credibility faster than a concession nobody expected.

Kind two: a real risk you control. Juniors losing the review and mentoring that builds expertise. This one isn’t about the industry, it’s about your team, which means your working agreements decide whether it comes true. The answer is a rule you set together, not a rebuttal.

Kind three: a misconception. “Every prompt is like pouring out a bottle of water.” The per-prompt numbers are knowable, and they don’t say that. Correct the arithmetic gently, with sources, and without dismissing the worry underneath it. The worry usually points at something real at a different scale.

Kind four: a value. “I didn’t get into this craft to supervise a machine.” You can’t argue someone out of a value, and trying reads as disrespect. What you can do is design roles so the value survives, and be honest about where the boundary sits.

One subtlety makes the whole sort work: the same concern can live in two kinds at once, at different scales. The environmental objection is true at industry scale and misread at the single-prompt scale. The sort forces you to say which scale you’re answering, out loud. Sorting is your actual job in this conversation; the debate is optional.

Here’s what the sort sounds like in an actual 1:1, in three moves.

First, ask for the strongest version: “Give me the strongest version of this worry. Convince me.” That one question replaces the debate with respect, and it surfaces the real concern instead of the polite one.

Second, name the kind out loud in plain words: “I think you’re right about this one,” or “this one is ours to fix,” or “I think the numbers say otherwise, let me show you,” or “that sounds like a value, and I won’t argue with a value.”

Third, before the conversation ends, commit to one follow-up in writing: the rule you’ll add, the number you’ll check, the thing that goes out-by-default. A sort that doesn’t end in a written commitment is just a nicer way of nodding.

So let’s put the framework to work, one conversation at a time.

Answer the seven concerns, one by one

The map below covers the seven objections I keep hearing. Six of them the talk maps better than anything else I’ve found; the seventh, what AI does to our thinking, has grown loud enough since to earn its own place. Your team may add others: discrimination baked into training data, deepfakes, privacy. The sort handles those the same way.

The planet

Give this concern its strongest version, with numbers that survive checking, because the talk’s own figures were compressed in places. The IEA’s 2025 Energy and AI report projects data-center electricity demand will more than double by 2030, to around 945 TWh, slightly more than Japan’s entire consumption today. Alex de Vries projected in Joule that AI-specific consumption alone could reach the scale of the Netherlands or Argentina by 2027. Water follows the same curve: Microsoft’s use jumped a third in 2022 while Google’s rose a fifth, and researchers behind the Making AI Less “Thirsty” paper project AI water withdrawal reaching half of the UK’s annual total by 2027. The verdict: at industry scale, your skeptic is right.

There’s a sharper twist worth conceding too. A Guardian analysis found the real emissions of the big providers’ own data centers ran about 7.6 times higher than officially reported. The trick is accounting: renewable-energy certificates let a company report clean power it never actually consumed at the site. Notice what that is: not a technology problem, an honesty problem.

Now the misread half. Per-prompt costs are measurable, and they’re small. Google’s published figure for a median Gemini text prompt: 0.24 Wh of energy and 0.26 mL of water, counting on-site water only. Mistral’s lifecycle study counts everything upstream, training included, and lands at 45 mL per response. The gap is scope, not virtue, and either way one prompt sits far below one kilometer of driving or one beef meal. Correcting this arithmetic matters, because a team member who quietly believes each prompt burns a lake will never use the tools well. The obvious comeback is fair: small times billions of prompts equals the buildout we just conceded. Both things are true. Your prompt is cheap; the trajectory is not; your subscription is a vote for it.

The move is to stop pretending it’s free, and to put that in writing. Three rules are worth adding to your team’s working agreements.

Default to the small model, escalate on need: most tasks don’t need the frontier model, and the price gap is the energy gap.

No AI where a script does the job: a regex, a spreadsheet formula, or a cron job costs nothing and never hallucinates.

And one screening question before any new AI use ships: what is this for, and would we defend it at scale?

Expect efficiency gains to get eaten unless you watch for them; the talk calls this the rebound effect, where every drop in cost per task invites more tasks.

When the conversation deserves more than a meeting slot, forward Andy Masley’s cheat sheet on AI and the environment. It’s the single best document I know for exactly this conversation: it runs the per-prompt arithmetic against everyday life (about 1,000 prompts to move your daily energy use by 1%) and still concedes the industry-scale buildout is a real question. Handing your skeptic the strongest version of the counter-arithmetic beats paraphrasing it in a meeting. What you can say: “You’re right about the trajectory. Here’s what our use actually costs, and here’s what we choose not to use it for.”

That answer covers the machines. The people who made the models possible are a harder conversation.

The people behind the data

The concern: the models were trained on work scraped from people who never consented, and the labor that made them safe was outsourced at poverty wages. The record backs it. TIME’s investigation documented Kenyan workers labeling toxic content for less than $2 an hour. And the talk’s own example has a scholarly source: Le Ludec, Cornet and Casilli documented French AI firms outsourcing annotation to Madagascar at poverty wages. This isn’t someone else’s supply chain. The verdict: they’re right, and no arithmetic rescues it.

So do

[truncated for AI cost control]