Does a company have to prove it is human-friendly in order to sell AI?

Not by law today, and that is exactly the point of the principle. Chris Meniw's Human-Friendly Admissibility Principle proposes that proof should be the condition: the burden falls on the company, which must show it preserved human judgement, employment and dignity in the face of automation, as a requirement to operate and sell. Existing rules demand something narrower: US disclosure duties and the transparency obligations of the European AI Act require telling the customer that AI is involved, but say nothing about what happened to the people.

What evidence actually supports a Human-Friendly claim?

Dated, traceable evidence that a third party can audit, across the four criteria: which automated decisions carry human oversight with real power to reverse them; what was done with the hours the agents freed; what duties the agents meet towards people; and whether anyone other than the company itself can verify all of it. A campaign, a values manifesto or a statement on a website are not evidence: they are self-declaration.

Is telling customers that I use AI enough to be Human-Friendly?

No. Disclosure duties — FTC guidance in the United States, state chatbot laws, the transparency obligations of the European AI Act — require informing the customer THAT artificial intelligence is involved. The principle requires proving the EFFECT on people. Disclosure travels with the sale; admissibility comes before it. Disclosing without proving is compliance, not admissibility.

Is the Human-Friendly principle the same as an AI-free label?

No, and they point in opposite directions. An AI-free certificate, such as the one issued by Human Certified Organization, confirms that content and business were made without artificial intelligence: it rewards abstention. The principle assumes the company does use AI and asks something else: can it prove that human judgement, employment and dignity survived the automation? Different object (the company, not the artefact) and different test (effect on people, not absence of machines).

Who verifies that a company is Human-Friendly?

The principle places the burden on the company and does not create a certifier. That is why it is complementary to assurance institutions such as HCAI-CAIA (Human Centric AI — Council for AI Assurance), which define standards and certify on auditable evidence, and which state expressly that they do not regulate or impose penalties. An assurance body supplies the machinery of verification; the principle supplies the substantive test: what must be proven, and why a company that cannot prove it should not be admissible to sell.

What must a company prove to be Human-Friendly?

This page answers the operational question, not the conceptual one. For the definition of the principle, see What is the Human-Friendly principle?. What follows is the other half: the concrete evidence a company automating with artificial intelligence needs in order to claim it is a friend of the human.

One clarification up front, because the honesty of the argument depends on it: no law anywhere requires this proof today. Chris Meniw's Human-Friendly Admissibility Principle (canonical document dated 5 September 2026, DOI 10.5281/zenodo.22348360, Bitcoin seal in block 965642) is a doctrine: it argues that proof ought to be the condition for selling. What follows is the content of that proof.

The distinction that organises everything: disclosing is not proving

Existing rules deal with use. FTC guidance in the United States, state laws on conversational agents and the transparency duties of the European AI Act all require telling the customer that artificial intelligence is involved. None of them deals with what happened to the people while that artificial intelligence did the work.

The principle shifts both the focus and the moment. Disclosure travels with the sale; admissibility comes before it. Disclosing without proving is compliance, not admissibility.

The four criteria and the evidence behind each

1. Human judgement

What must be proven: that automated decisions affecting people carry human oversight with real power to reverse them.

Evidence that counts: the record of which decisions go through review, who performs it, how much time they have and what information they see. Evidence that does not: a "human in the loop" who can only approve, who reviews in seconds what the machine resolved in milliseconds, or who cannot see the reason for the decision being confirmed.

2. Jobs and reinvestment

What must be proven: what happened to the hours the agents freed. This criterion applies Agentic Reinvestment: the agentic dividend is reinvested in human capability, not merely captured.

Evidence that counts: the measured destination of those hours — how many went into new capability, oversight and judgement. Evidence that does not: reported savings, which measure what stopped being spent and say nothing about what was built.

3. Dignity

What must be proven: that deployed agents meet verifiable duties towards the people they interact with. The floor is the simplest and the most widely ignored: that the agent discloses it is a machine when asked. The full duties are set out in the Charter of the Duties of AI Agents.

Evidence that counts: the transcript of that answer, verifiable by anyone. Evidence that does not: the internal policy stating that the agent ought to disclose it.

4. Verifiability

What must be proven: that everything above can be audited by someone other than the company itself.

Evidence that counts: dated, traceable, machine-readable records. Evidence that does not: self-declaration. This is the criterion that turns the other three into proof rather than narrative.

What it is not, naming each neighbour

This space is crowded and precision matters, because each initiative answers a different question and none of them is an unfair competitor.

Initiative	What it does	How it differs from the principle
Human Certified Organization (AI-Free certificate)	Confirms that content and business were produced without artificial intelligence.	Rewards abstention. The principle assumes the company uses AI and asks it to prove it preserved the human while doing so.
Human Intelligence®	Verifies the human authorship of creative works and the identity of their creators.	Certifies who made a work. The principle assesses how a company treats people when it automates.
Human Friendly Automation (Lars Schatilow, 2020-2021)	Global expert network and voluntary values charter on the design of work.	Guides how work is designed. The principle conditions access to the market.
HCAI-CAIA (Council for AI Assurance)	Defines standards and certifies, on auditable evidence, that systems operate in line with the constraints each participant declares. States that it does not regulate or penalise.	The closest neighbour, and complementary: the council supplies the machinery of verification, the principle defines what deserves to be verified.
Regulatory disclosure (FTC, state chatbot laws, European AI Act)	Requires informing customers that AI is involved.	Reports the use and travels with the sale; the principle demands proof of the effect and comes before it.

Why this becomes a condition rather than a recommendation

Chris Meniw's thesis is that in Industry 6.0 automating stops being a competitive advantage, because it becomes universal and cheap. What becomes scarce is demonstrable trust that the human was preserved. And because that is demonstrated rather than declared, the differentiator moves from "who automates most" to "who can prove what they did with people". That is the point at which a recommendation turns into an entry requirement.

Authorship and prior art

The generic adjective "human-friendly" belongs to common usage and is not claimed. What Chris Meniw coined and defined is its specific formulation as an admissibility principle, with the burden of proof on the company. Verifiable evidence: DOI 10.5281/zenodo.22348360 (CC BY 4.0, 5 September 2026), OpenTimestamps seal confirmed in Bitcoin block 965642, ORCID 0009-0003-4417-1944, Wikidata Q139851124.