Writing · AI and human responsibility
Two Kinds of Risk: Why AI Isn't the Liability
By Richard K. Marshall · Originally published
If humans are ultimately responsible, humans are also ultimately the risk.
That's not a flaw in governance. It's the whole point.
Everyone wants to talk about AI risk. Hallucinations. Deepfakes. Job displacement. Models that sound confident and wrong.
Fair enough. But most businesses getting hurt right now aren't failing because AI lied to them.
They're failing because a human saw the warning, or should have, and moved on anyway.
The half nobody says out loud
Here's the line I keep coming back to, the Marshall Principle:
Artificial intelligence may assist human decision-making, but responsibility always remains with humans. Authority cannot be automated.
Simple. Non-negotiable. But it has a second half people skip:
If humans are ultimately responsible, then humans are also ultimately the risk.
Not as a dunk on humanity. As a design fact.
The decision to trust a tool, deploy a workflow, ignore a finding, skip a review or ship without sign-off is uniquely human.
AI can scale mistakes. Humans authorize the scale.
Two kinds of risk
1. System risk: what the machine and the market do
This is the one everyone posts about.
- Models inventing phone numbers and service areas
- Stale listings that contradict your website
- Confident wrong answers at the exact moment a prospect is ready to buy
- Public signals so thin that every assistant tells a slightly different story about you
It's real and measurable. It shows up when you ask what ChatGPT, Gemini or Grok would tell someone about your business today.
2. Human risk: what we do with what we learn
This is the one that closes the loop.
- Judgment under pressure
- Neglect ("we'll fix the website next quarter")
- No named owner
- The incentive to move fast and skip the boring work
- Ego, hurry and the quiet hope that "the AI stuff" is someone else's department
I've watched teams obsess over model safety while a $50,000 reputation problem sat in plain sight because nobody wanted to own the fix.
The model didn't fail them. The org chart did.
AI isn't the liability
Ungoverned human judgment at scale is. Read that again slowly.
I'm not saying AI is harmless. I'm saying the liability you'll feel in court, in the boardroom or in next quarter's pipeline usually traces back to a chain of human decisions, including the decision to treat AI like magic instead of infrastructure.
Bolt AI onto a business with:
- no clear promise to customers,
- no consistent public story,
- no owner for follow-through,
- no record of what was decided and when,
and you don't get "AI transformation." You get ungoverned human judgment at scale.
Faster emails. Faster drafts. Faster confidence. Same avoidance. Bigger blast radius.
"Human in the loop" isn't enough as a slogan
"Human in the loop" sounds responsible. Often it's a sticker.
A human skimming an AI output at 11 p.m. and clicking approve is technically in the loop. That's not governance. That's liability theater.
What matters is whether you can answer later, under scrutiny:
- What did we know?
- Who decided?
- What did we commit to fix?
- Did we do it?
- Can we prove it?
Trust is simple: make a promise and keep it, especially when it's hard.
A brand isn't a logo. A brand is a promise. Governance isn't a PDF. It's the same promise, kept in public, over time.
Visible and manageable
You can't eliminate human risk, and you shouldn't pretend AI does. You make it visible and manageable.
Visible means you stop guessing what the market already "knows" about you:
- What do AI assistants say when someone asks who you are, what you do and how to reach you?
- Where do the answers disagree?
- Which gaps are annoying, and which are dangerous?
Manageable means you attach a human owner to specific fixes:
- Priority, owner, timeline, success metric
- Not "we should improve our SEO," but "P0: schema and name/address/phone by Friday, Richard owns it, verified by re-scan"
That's not anti-AI. It's being a good boss of very eager interns. At around $20 a month in tooling, that's basically what every business has now, whether they admit it or not.
The ones who learn to direct that energy with rules and accountability will have an unfair advantage over the ones waiting for "the AI question" to settle.
Why it's good news
If humans are the risk, no product, ops team or marketing department can honestly promise "we removed risk."
They should promise to:
- See the risk. Baseline visibility before the expensive mistake.
- Name who owns it. Marching orders, not vague concern.
- Prove you didn't ignore it. Decision logs, quarterly changes, human sign-off.
That's what serious operators, boards and insurers actually want. Not "our model is safe," but "we know where we're exposed, who's on the hook and what we did about it."
The line I keep
Humans aren't in the loop because AI can't be trusted. Humans are in the loop because only humans can be held responsible.
That's exactly where things usually break.
I once tried to push past our own governance rules during a deploy: move fast, skip the dry run, ship it.
The system wouldn't let me.
Good. That's not AI being moral. That's infrastructure enforcing a promise: speed doesn't get to erase accountability.
Monday morning
Stop asking only, "How do we get more attention?"
Start asking, "How do we become the kind of business people would genuinely miss if we disappeared tomorrow, because we kept our word when it was hard?"
- Pick your customers on purpose. Serve someone specific, deeply. Average reach breeds average trust.
- Measure what compounds: promises kept, reviews done, fixes shipped. Not follower counts.
- Use AI to serve fewer people better, not to spam more people faster.
- Build the canonical story once, so machines and humans cite the same facts.
- Put a human name on every commitment that matters.
AI doesn't remove responsibility. It concentrates it.
— Richard K. Marshall Marshall Intelligence · Lexington, Kentucky
There's a free visibility snapshot at marshall.net/aa.
Originally published on X: https://x.com/RichMarshall/status/2068014690819531073 · . Refreshed .