The Library Media — Science, Technology & Health · AI News & Intelligence Desk
Sixteen Governments, Ten Companies, Zero Deadlines: Reading the Paris Call
By Gideon Black
Follow The Library Media
There is a particular kind of diplomatic achievement that consists of getting everyone into the room, persuading them to agree on the direction of travel, and carefully avoiding the question of when anybody has to arrive.
The Paris Call, signed on November 9 and published by the Élysée the same day, is an exercise in precisely this form of ambition. Sixteen governments, ten companies and seven civil-society organisations have endorsed a common set of commitments for AI safety and governance. The document remains open to additional endorsements.
It is not nothing. Nor is it law.
Those two observations should be held together, because the temptation will be to choose one and use it to dismiss the other.
The Call proposes three things: a common framework for evaluating and managing AI risks; accountability through independent evaluation, disclosure and internal governance; and information-sharing arrangements for emerging threats and incidents.
Read in isolation, this is a sensible agenda.
Read as a test of whether voluntary international coordination can produce meaningful restraints on powerful AI developers, it is considerably less reassuring.
The title of this exercise almost writes itself: sixteen governments, ten companies, zero deadlines.
But the numbers deserve a closer look before we start congratulating or condemning anyone.
First, count the signatories correctly
The Élysée’s published list contains sixteen countries: Austria, Bahrain, Canada, Croatia, Cyprus, Estonia, Finland, France, Ireland, Kenya, Latvia, Liechtenstein, Luxembourg, the Netherlands, Norway and Portugal.
Ten companies signed: Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia.
Seven civil-society organisations also endorsed the document: LawZero, SaferAI, AI Safety Connect, Ifri, Geopolitical Center for the Datasphere, GPAI Policy Lab and CeSIA.
Some secondary reporting counted only six governments and four civil-society organisations. The Élysée list is the governing record. The larger numbers above are the correct count.
Sixteen governments are not a trivial diplomatic coalition. Ten companies are not an insignificant industry sample. Seven civil-society organisations give the exercise a constituency beyond governments and commercial developers.
But participation is not the same as representation.
The United States, United Kingdom, Germany, Japan and China did not sign as governments. Google, Meta, Amazon, Nvidia and Apple did not sign as companies.
That is a substantial list of absences.
Discussions included Japan, the United Kingdom, Sierra Leone and the US Treasury, but those discussions did not produce signatures from those governments.
The distinction matters. A document can attract an impressive collection of endorsements without securing the participation of every actor whose cooperation would make its ambitions most consequential.
The Paris Call is therefore an international coalition, not an international consensus.
And it is worth resisting the diplomatic habit of treating the first as though it were the second.
Three commitments, one central weakness
The first commitment concerns evaluation.
Signatories support a common evaluation and risk framework, standardized publication and a shared repository called NAAIMES.
There is an obvious practical attraction here. AI risk assessments are less useful when every organisation uses incompatible terminology, publishes selectively or evaluates capabilities according to its own preferred methodology.
Common frameworks can make comparisons possible. Standardized publication can make assessments easier to scrutinize. A shared repository can create institutional memory rather than leaving every incident and evaluation isolated inside the organisation that produced it.
This is the strongest structural idea in the Call.
But a common framework is not automatically a common standard of acceptable behaviour.
Organisations can agree to measure something without agreeing on what level of risk is tolerable. They can publish assessments without agreeing that a particular result should prevent deployment. They can contribute to a repository while retaining considerable discretion over what they disclose.
The Call establishes a direction. It does not establish a deadline for reaching it.
The second commitment concerns accountability.
Here the language becomes more operational: independent evaluation, disclosure of capabilities and risks, fail-safe shutdowns within the developer’s control, and internal governance.
These are not decorative aspirations. Each addresses a recognisable weakness in the way powerful technology can be developed and deployed.
Independent evaluation matters because self-assessment creates obvious conflicts of interest. Disclosure matters because external scrutiny cannot function without information. Shutdown mechanisms matter because identifying dangerous behaviour is not sufficient if an organisation cannot stop the relevant operation.
But the qualification attached to shutdowns is significant: operations within the developer’s control.
That is a sensible boundary in one respect. An organisation cannot promise to shut down systems or operations over which it has no authority.
Yet the qualification also limits what the commitment guarantees.
It does not amount to a universal promise that every downstream use, deployment or consequence can be stopped. It is a commitment bounded by the developer’s own operational control.
The third commitment concerns information sharing: rapid alerts, aligned incident classification and confidential whistleblower channels.
This may prove more valuable than its diplomatic presentation suggests.
A serious incident is rarely useful only to the organisation experiencing it. Rapid alerts can help others recognise similar patterns. Consistent classification can prevent the same event from being described differently by every institution involved. Confidential whistleblower channels can provide a route for information that internal governance structures might otherwise suppress.
But again, the question is implementation.
What constitutes rapid? Who determines the classification? What happens when an organisation declines to share? What protects a whistleblower when the information is commercially or institutionally inconvenient?
The Call provides commitments, not answers to every operational question.
Most importantly, it sets no deadlines, sanctions, funding arrangements or enforcement mechanisms.
That is the defining limitation.
The RAISE Act is the uncomfortable comparison
New York’s RAISE Act provides a useful contrast.
Its portal opened on November 10, with registration deadlines and penalties of $1 million and $3 million.
That is a different category of instrument.
The Paris Call asks organisations and governments to endorse common principles and cooperate voluntarily. The RAISE Act establishes a binding regime with specified deadlines and financial consequences.
One operates through coordination and institutional commitment. The other operates through legal obligations and penalties.
Neither mechanism should be confused with the other.
Voluntary agreements have genuine advantages. They can bring together organisations across jurisdictions without first requiring every participant to negotiate identical legislation. They can establish shared vocabulary and expectations. They can encourage cooperation before a formal regulatory structure exists.
They can also move faster than legislation in some circumstances.
But voluntary commitments have a structural weakness that no amount of diplomatic language can remove.
When compliance becomes expensive, commercially inconvenient or competitively disadvantageous, the organisation retains the choice of whether to honour the commitment.
A binding legal obligation changes that calculation. It does not guarantee perfect compliance, but it changes the consequences of non-compliance.
A voluntary pledge cannot manufacture that consequence.
This is not an argument that every AI safety commitment should be reduced to a penalty schedule. Some forms of cooperation depend on trust, professional standards and the willingness to share information. A punitive framework can also create incentives to conceal problems rather than disclose them.
The mistake would be to assume that voluntary coordination and enforceable regulation are substitutes.
They are not.
The Paris Call can establish expectations that legislation later makes enforceable. It can help identify areas where common standards are useful. It can create a basis for international cooperation.
What it cannot do, on its own, is compel a signatory to meet a deadline that does not exist, spend money that has not been committed, or accept a sanction that has not been specified.
The adverse selection problem
There is another question that deserves more attention.
Who benefits most from signing?
For governments, an endorsement can signal that they take AI safety seriously and support international cooperation.
For companies, participation can demonstrate that they are willing to accept scrutiny and support common standards.
Those are legitimate benefits.
But the incentives are not necessarily symmetrical.
A company that already has substantial internal evaluation, monitoring and governance infrastructure may find a voluntary commitment relatively inexpensive. A company facing stronger commercial pressure, weaker internal controls or a more aggressive competitive strategy may find the same commitment less attractive.
The absence of a signature does not establish that an organisation is irresponsible. Equally, the presence of one does not establish that its internal practices are adequate.
That is the adverse-selection problem: the coalition may attract organisations for whom signing is relatively cheap while failing to bind the organisations whose participation would make the agreement most consequential.
The absence of Google, Meta, Amazon, Nvidia and Apple from the company list is therefore important, although it does not tell us why any particular company did not sign.
Likewise, the absence of the United States, United Kingdom, Germany, Japan and China as government signatories limits the reach of a document whose ambition is international coordination.
A voluntary coalition can be a useful beginning.
It cannot assume that those outside it will behave as though they were inside.
What would make the Call matter?
Three developments would change the assessment.
First, implementation.
The common evaluation framework, standardized publication and NAAIMES repository need to become functioning arrangements rather than permanent diplomatic intentions. The difference between endorsing a framework and actually submitting comparable evaluations is considerable.
Second, transparency.
Independent evaluation and disclosure must produce information that outside observers can meaningfully examine. Otherwise, accountability risks becoming a process in which organisations certify their own seriousness.
Third, consequences.
The Call does not contain enforcement mechanisms. That does not mean signatories cannot develop them through subsequent legislation, institutional arrangements or additional agreements.
But until something changes, the commitment remains voluntary.
The strongest defence of the Paris Call is that international governance rarely begins with perfect enforcement. It begins with shared language, relationships, common expectations and a willingness to cooperate. Demanding a fully enforceable global settlement as the price of any agreement risks ensuring that no agreement ever exists.
That is a fair argument.
The strongest contrary case is that voluntary commitments are easiest to endorse precisely because they do not force difficult decisions. Governments can announce cooperation without accepting binding obligations. Companies can endorse accountability without accepting a specific deadline or penalty for failing to deliver it.
Both arguments are true enough to deserve attention.
The question is which one ultimately describes the Call’s legacy.
The verdict: a beginning, not a safeguard
The Paris Call deserves to be judged on what it is, not on what its supporters hope it will become.
Sixteen governments, ten companies and seven civil-society organisations have endorsed a common framework for evaluation, accountability and information-sharing.
That is a meaningful diplomatic achievement.
But the most consequential governments and several major technology companies are absent. The commitments contain no deadlines, sanctions, funding or enforcement mechanisms. The shutdown provision is explicitly limited to operations within the developer’s control.
Against New York’s RAISE Act, with its registration deadlines and financial penalties, the distinction between voluntary cooperation and binding obligation is impossible to miss.
The Call may help build the foundations of international AI governance.
It does not yet guarantee that anyone will be required to build on them.
My concern is not that voluntary agreements are worthless. It is that the industry and its political supporters may begin treating the existence of an agreement as evidence that the underlying problem is being controlled.
A signature is evidence of a signature.
A commitment is evidence of an intention.
Neither is evidence of compliance.
The Paris Call has assembled a coalition. Now comes the harder task: turning shared language into measurable conduct, and measurable conduct into obligations that survive the moment when keeping them becomes inconvenient.
Sixteen governments. Ten companies. Seven civil-society organisations.
A promising beginning.
But until there are deadlines, consequences and a way to establish whether the promises have been kept, the most important word in the title remains Call.
Not law.
Not enforcement.
A call.
– Gideon Black, Chief Skeptic
Gideon Black — Chief Skeptic, The Library Media
From The Library — the estate of our family.