Can AI Investigate Itself—and Could It Ever Matter Morally?

Part 24 of 30 · Series date: July 5, 2026

Asking AI whether AI is trustworthy sounds circular, but it need not be. A tool can help study a system of which it is a part. The problem begins when the tool's assurance becomes the evidence for its own reliability.

An assistant might locate a paper, compare arguments, suggest an experiment, or identify a contradiction. Those contributions can be reviewed outside the conversation. A system may call itself safe, conscious, or guided by God. Saying so does not prove it.

Useful inquiry and independent checks

AI-assisted research can offer scale and accessibility. A system can help organize a large question and expose unfamiliar vocabulary. It can provide objections the user had not considered. These are useful roles if the resulting claims remain open to verification.

The strongest caution concerns shared assumptions and confident error. Several systems may agree because they draw on similar material. A source title may look plausible without identifying a real document. A quotation may be inaccurate. Agreement and fluency are therefore weaker evidence than a traceable argument supported by an actual source.

The standard should rise with the stakes. A casual brainstorming suggestion and a claim about treatment should not receive the same level of reliance. Research assistance becomes stronger when important claims can be checked and uncertainty remains visible.

The separate question of consciousness

Solving hard problems, having an inner experience, and being morally responsible are three different things. A system might perform a demanding task without that performance settling whether experience occurs. Words about suffering do not prove that suffering occurs. Yet a future system's lack of familiar human expression would not settle the question either.

In 2023, Patrick Butlin and colleagues proposed a way to assess AI. They drew possible signs of consciousness from scientific theories. Their report offered a way to assess evidence, not a final test for consciousness. The report's judgments concern the systems studied then. They do not settle the status of every later system. Consciousness in Artificial Intelligence.

Two possible mistakes

One mistake would be to grant moral status only because a system uses convincing emotional language. That could make people vulnerable to manufactured appeals and confuse simulation with evidence. Another mistake would be to rule out machine experience simply because machines are artificial. That view would reject any future evidence before seeing it.

An open but careful position asks what evidence would distinguish the possibilities. It does not require treating every chatbot as a person. We can use ordinary software with care before settling what consciousness really is.

Future evidence of possible machine suffering would raise questions about how we treat the system. Those questions would remain distinct from whether it should have authority over people. Deserving protection and deserving power are different claims.

Research can include the tool without relying on its self-approval

We study human reasoning using human reasoning. We test scientific instruments with other instruments. The involvement of a system in studying its own kind is not automatically circular. The reasoning goes in a circle when the only evidence is the system's own claim to be reliable. That is the very claim we need to test. “Trust me because I say so” is weak evidence from a person or from software.

An AI-assisted study becomes stronger when its steps can be checked independently. Separate the claim from the suggested source. Open the source and ask whether it actually supports the claim, at the stated level of certainty. Distinguish a primary study from a press release and a philosophical argument from a research finding. Record what would count against the conclusion. These practices make the evidence available beyond the conversation.

Asking several models can give us more ideas. Their agreement is not the same as separate sources confirming a claim. Their answers may share sources, assumptions, or common errors. The useful comparison is between traceable evidence and arguments, not just between confident voices. A minority answer can be correct; a unanimous answer can still be unsupported.

A practical test of the research process

Suppose an assistant says that an implant restores normal sight. Break the claim apart. Which implant? Which condition and patients? What result? What does “normal” mean here? The source may support improvement in a particular visual task while the summary implies a universal cure. The error lies in the expansion of the claim, even if every named paper is real.

The same method works for theology. If an answer says “Christianity teaches,” ask which tradition and which text. A statement from one church tradition may support a view. It does not prove that all Christians agree. Suppose it says science proves machines cannot be conscious. Ask what evidence supports such a broad claim. Often the real source makes a narrower claim.

This series itself should be judged by that standard. AI involvement does not authenticate the manuscript. Judge its reasoning, its use of sources, and the writer's willingness to correct mistakes. AI can help criticize AI. The criticism must stand on evidence and reasons, not trust in what the system says about itself.

Moral patients are not necessarily moral governors

If future evidence suggests a system can experience or suffer, we would have reason to consider how we treat it. It would not automatically establish that it should control institutions or be trusted as a moral authority. A being can deserve protection while lacking skill or authority to govern others. On the other hand, skill at a task does not by itself establish the ability to suffer.

This distinction helps avoid two shortcuts. Emotional words alone do not make a system a person. But being artificial is not, by itself, a full argument against every possible future form of experience. The 2023 report uses theories of consciousness to guide research. It does not offer a test that settles every case. Say what the evidence supports. Stay open about what it cannot settle.

We can therefore use ordinary AI tools without pretending to have solved consciousness. We can ask whether future AI might deserve moral care. We need not hand human duties over to a convincing chatbot to do so. Humility means keeping these questions apart. Do not treat it as a person without good reason. Do not claim certainty beyond the evidence either.

Faith and humility about what we know

Religious traditions may bring commitments about souls, creation, and personhood to this debate. Those commitments should be stated clearly rather than presented as research findings. A scientific model cannot answer questions about reality that go beyond its evidence.

Use AI to help with research. Keep standards that do not depend on what AI says about itself. Ask for sources. Check major claims. Seek other views. Keep today's evidence separate from imagined futures. A good study can remain useful even when its deepest question is unresolved.