When anything can be generated, the evidence might matter as much as the artefact.
For a long time, institutions treated the finished artefact as the primary object of trust. A report on official letterhead, an invoice, a photograph, a contract, a set of accounts or a published article carried authority because producing it required time, expertise and access to particular tools. The form itself was evidence of effort. The effort, in turn, was taken as a rough proxy for reliability.
Generative AI unsettles that arrangement.
A well-written briefing can now be produced in seconds. So can a plausible image of an event that did not occur, an audio recording of someone who said nothing, a convincing-looking data visualisation, a legal letter, a student assignment or a board paper. The ability to generate a persuasive object is no longer scarce. It is increasingly available to anyone with a modest level of technical capability.
This does not mean that truth has become impossible, nor that every digital artefact should be met with suspicion. It means that the basis of trust is changing. When the artefact can be produced cheaply and at scale, attention shifts to its provenance. Who made it? Under what authority? From what evidence? What process produced it? Who will stand behind it when it matters?
The question is not simply whether something is real. It is: who vouches for it?
The declining authority of appearance
This is not the first moment in which technology has challenged established forms of evidence. Photography changed the status of visual representation. Digital editing made photography easier to manipulate. The internet disrupted the gatekeeping power of publishers, broadcasters and professional bodies. Social media made distribution cheap and immediate.
Generative systems differ because they compress the distance between intention and artefact. A person no longer needs to write the report, compose the image, record the voice or build the presentation in order to produce something that resembles those things. The resulting object may be useful, accurate and legitimate. It may also be synthetic, misleading or assembled from material whose origin is difficult to establish.
The point is not that generated material is inherently less valuable. Some of it will be excellent. Organisations will use it to reduce routine workload, improve communication, support accessibility and make specialist expertise more widely available. The problem is that appearance can no longer carry the evidentiary weight it once did.
A polished document does not tell us whether its claims have been checked. A confident voice does not tell us whether the speaker exists. A realistic image does not tell us what happened. The aesthetic signals that once helped people make quick judgements about credibility are becoming weaker.
Trust must therefore become more explicit.
From authenticity to accountability
Authenticity is often treated as the central challenge of the generative era. There will be important technical work on watermarking, content credentials, cryptographic signatures and tamper-evident records. These tools matter, particularly for journalism, public administration, financial transactions, legal documents and elections.
But authenticity alone is not enough.
A document may be authentic in the narrow sense that we can identify the system or account that produced it. That does not establish whether it is accurate, fair, authorised or fit for purpose. A report generated from a verified organisational account can still contain weak analysis. A genuine image can still be used deceptively. A real person can still make an irresponsible claim.
The deeper issue is accountability.
A receipt matters because it links an artefact to a chain of responsibility. It tells us not only that something exists, but where it came from, what was exchanged, when the exchange occurred and who is accountable for the transaction. It makes a claim about provenance that can be checked.
In organisational settings, a comparable record should accompany important AI-generated material. Not necessarily a burdensome technical log for every email or draft paragraph, but a proportionate account of authorship, evidence, review and authority. For a policy recommendation, this may include the source material used, the analytical assumptions, the people who reviewed it and the official who approved it. For a public-facing image, it may mean a clear statement that it is synthetic. For a financial or employment decision, it requires a traceable record of how the decision was reached and who had authority to make it.
The relevant test is practical: if this artefact is challenged, can the organisation explain its origin and defend its use?
Vouching is an institutional function
To vouch for something is more than to say that one believes it. It is to put one’s own standing behind a claim. In professional life, this is a familiar responsibility. An auditor signs an audit opinion. A clinician documents a diagnosis. A public servant briefs a minister. A researcher identifies methods and sources. A director accepts obligations attached to a board decision.
These acts do not guarantee that the work is correct. They make error visible and responsibility identifiable. They create incentives for care because someone has a reason to take the work seriously before it is released.
Generative AI can weaken this discipline if organisations allow systems to become a convenient source of plausible but ownerless output. A manager may circulate an AI-generated analysis without knowing its evidentiary basis. A staff member may rely on a synthetic summary instead of reading the underlying material. An executive may approve a recommendation on the assumption that the system has undertaken a form of analysis which it has not, in fact, performed.
In such cases, the technology does not simply create a risk of factual error. It diffuses responsibility. The artefact appears without a clear author, reviewer or accountable decision-maker. When it fails, every participant can point elsewhere: to the software provider, the user, the data, the prompt, the person who approved it or the process that did not require further checking.
That is not a governance model. It is an abdication of governance.
The responsibility for an organisational decision cannot be assigned to a model. Systems can provide information, generate options and assist with analysis. They cannot hold a professional duty, respond to a constituent, explain a trade-off in a public forum or accept the consequences of a decision.
People and institutions must still do that work.
Provenance as a public good
The value of provenance extends beyond risk management. It is a condition of a functioning information environment.
Markets depend on credible signals. Citizens depend on reliable public information. Researchers depend on traceable evidence. Workers need to know whether a communication comes from their employer, a colleague or an automated system acting in someone’s name. Consumers need to know whether an endorsement is genuine, generated or paid for.
When credible provenance is absent, people incur costs to verify basic claims. They spend more time checking sources, consulting alternative accounts and deciding whom to trust. The burden does not fall evenly. Well-resourced organisations can employ experts, legal advisers and verification tools. Individuals and smaller organisations may not have the same capacity. A high-noise information environment therefore has distributive consequences.
There is also a collective cost. If fabricated material becomes common enough, authentic material may lose its persuasive force. A person who is recorded making a damaging statement may claim, plausibly enough, that the recording is synthetic. A genuine image of abuse or wrongdoing can be dismissed as generated. The problem is not simply false content being believed. It is true content becoming easier to deny.
This is sometimes described as the “liar’s dividend”. The phrase is useful because it points to the political character of the problem. Doubt can be strategically manufactured. Those with power may benefit from an environment in which evidence is endlessly contestable and responsibility is difficult to locate.
The response cannot rest solely on better detection systems. It must involve institutions that can credibly vouch for information and explain the basis on which they do so.
The receipt is not bureaucracy for its own sake
There is a risk that calls for provenance become a demand for cumbersome compliance. That would be a mistake. Not every generated note, internal draft or routine communication needs a formal chain of custody. Organisations should not turn ordinary work into a procession of attestations.
The standard should be proportionate to consequence.
Where the content is low stakes and readily reversible, light disclosure may be sufficient. A note that a document was drafted with AI assistance, followed by human review, may be all that is needed. Where the material informs a consequential decision, the standard should rise. The relevant evidence must be available. Assumptions should be open to scrutiny. The responsible person should be identified. Where rights or significant interests are affected, people should have access to review and correction.
This approach has a further benefit. It directs attention away from abstract arguments about whether a system is intelligent and towards the practical question of whether the organisation’s decision process is sound.
A model may be highly capable, but capability does not substitute for authorisation. A generated analysis may be accurate, but accuracy does not settle whether the analysis was properly used. An automated recommendation may be efficient, but efficiency does not eliminate the need for an accountable decision-maker.
The receipt is a small but useful metaphor for this distinction. It says: here is the record, here are the parties, here is what occurred, and here is the basis on which the claim can be checked.
A new literacy of trust
The generative era will require a more developed public literacy about information. People will need to ask different questions of the material in front of them. Is this original reporting, commentary, synthesis or generation? What evidence does it rely upon? Is there a responsible author or institution? Can the claim be verified independently? What interest might sit behind its circulation?
Organisations need a comparable literacy. Staff should know when to disclose AI assistance, when to seek review, what sources are acceptable for a task and how to distinguish a useful draft from a defensible conclusion. Leaders should be able to identify where generated output enters significant processes and whether their governance arrangements still allocate responsibility clearly.
This is not an argument for nostalgia. The earlier information order was never as reliable or democratic as it is sometimes portrayed. Powerful institutions have always been able to circulate misleading material, and formal credentials have never been a perfect guarantee of truth. Generative technology may also reduce barriers to participation and allow more people to communicate effectively.
Yet its benefits will be weakened if organisations use it to remove the very signals that allow others to assess claims critically.
The future of trust will not depend on whether we can distinguish every synthetic artefact from every human-made one. That standard is neither realistic nor sufficient. It will depend on whether the institutions that make consequential claims are prepared to stand behind them.
When anything can be generated, the artefact is no longer enough. What matters is the record of responsibility attached to it, and the person or institution willing to say: this is ours, this is how it was made, and this is what we are prepared to answer for.
