Who is responsible when AI-assisted migration advice is wrong?

OMARA's guidance on AI does not leave the responsibility question open. It permits registered migration agents to use AI, then states in unqualified terms who carries the consequences when the output is wrong. The answer is the agent, and the regulator's reasoning is source-independence: where the information came from is irrelevant to who is answerable for it. That single phrase — “regardless of the source” — is the spine of the guidance, and the rest of the scheme points the same way.

The permission itself, and the written consent required before personal details go into a tool, are covered in our article on whether migration agents can use ChatGPT. This article is about the allocation of responsibility.

1. The regulator's answer, in full

“RMAs may choose to use AI while providing immigration assistance. However, RMAs maintain responsibility for the immigration assistance they give. This means that an RMA who uses AI and gives inadequate or inaccurate information to a consumer remains responsible for that information, regardless of the source.”

Three moves in three sentences. The first is permission: AI is something an RMA may choose to use. The second is the pivot — “However, RMAs maintain responsibility”. The third states the allocation and closes the argument: “regardless of the source”.

Two choices of wording carry the weight.

• Maintain is the verb of continuity: responsibility is not created by the use of AI, and it cannot be shed by it. It was there before the tool and it is still there afterwards.
• “regardless of the source” is the regulator anticipating the response that the error came from somewhere else, and rejecting it in advance. A hallucinated regulation, a wrong processing time, a mis-summarised policy instrument: once the agent gives that information to a consumer, it is the agent's information, and the agent remains responsible for it.

2. The chain behind the answer: sections 276 and 280 of the Act

The Act does not regulate software. It regulates a person who gives immigration assistance, and OMARA's summary of that scheme is direct:

“Under section 280 of the Migration Act 1958 (Cth) (the Act) only certain people can provide lawful immigration assistance in Australia. They are:”

The guidance lists them:

“registered migration agents (RMAs)”

“Australian legal practitioners”

“a small number of exempt persons such as close family members of a visa applicant.”

Every category is a natural person. A model cannot be registered as a migration agent, cannot be admitted as a legal practitioner, and cannot be a close family member. So a model can never be the lawful provider. When AI-assisted immigration assistance is given lawfully, it is given by a person who falls into one of those categories — and that person is the one the framework holds. What counts as assistance is defined separately, and it includes exactly the work an agent might hand to a model:

“Immigration assistance is defined under section 276 of the Act and includes a range of activities such as preparing a visa application.”

The guidance then closes the route in two steps, and applies them:

“Giving immigration assistance in Australia in circumstances other than that outlined in section 280 of the Act is unlawful.”

“There are no exemptions, including information generated through AI.”

“Depending on the circumstances, a person who provides immigration assistance in Australia through AI may still commit an offence under section 280 of the Act.”

Routing assistance through a model changes the drafting process. It does not add the model to the list of people who may lawfully provide assistance, and it does not remove the person using it from the list of people who are answerable. The question the Act asks is who provided the assistance. The answer is never the model.

3. Why the AI vendor does not take the error

The intuition is almost universal: the model generated the error, so the error belongs to whoever built the model — or at least it is shared. Under this framework that is wrong, and it is wrong in a specific way.

• The guidance places no obligation on an AI vendor or platform. The conduct it addresses — and the conduct a consumer can complain about — is “the conduct of an RMA”. Responsibility is attached to the person who gives the assistance, not to the provider of the tool they used to prepare it.
• Responsibility attaches to the information given to a consumer, not to the process that produced it. The operative words are “gives inadequate or inaccurate information to a consumer”. A draft the agent reviews and corrects never becomes information given to a consumer. A draft that goes out unexamined does, and it arrives carrying the agent's responsibility.
• Vendor terms cannot reallocate a statutory obligation. An AI provider's disclaimers describe that provider's relationship with its customer. They do not touch the agent's obligations under the Act or the Code, and OMARA's guidance makes the agent's responsibility conditional on nothing the vendor says.

There is no exemption anywhere in the guidance for information generated by a machine, and no provision that transfers the agent's responsibility to the machine's maker. The framework is source-blind by design: it looks at the assistance as given to the consumer, and at the person who gave it.

This article records what the guidance and the Act provide; it does not advise on any individual matter.

4. What a consumer can do, and what OMARA can do

“Consumers with concerns about the conduct of an RMA who has used AI to give immigration assistance can lodge a complaint with the OMARA.”

“If the OMARA has reasonable suspicions that an RMA’s conduct may have breached the Code, it can investigate the RMA and take disciplinary action where appropriate.”

Two things follow from that passage.

• The complaint is about “the conduct of an RMA”. The guidance routes the concern to the professional regulator of the agent — not to the provider of the tool.
• The investigatory threshold is “reasonable suspicions” that conduct “may have breached the Code”. Proof of a breach is not the trigger; a reasonable suspicion is enough for OMARA to investigate, and disciplinary action can follow where appropriate. Since the Code's standards apply to AI-assisted work as they do to any other work, the fact that a model produced the wrong answer does not change that analysis.

For a practice, this is the exposure that matters. It is not a hypothetical claim against a model provider; it is a complaint about the agent's own conduct, investigated by the body that controls the agent's registration.

What this requires in practice

• Treat every AI output as the agent's own draft. It becomes the agent's information at the moment it is given to a consumer.
• Put a named human review step between generation and delivery, and record who approved what. If OMARA ever asks, the file is what shows that responsibility was exercised rather than merely assumed.
• Never let the AI said it stand as an explanation — to a client, to a colleague, or to OMARA. The guidance contains no such defence.
• Correct errors that do reach a client. The guidance attaches responsibility to inadequate or inaccurate information given to a consumer, and putting that right is the agent's obligation, not the vendor's.
• Do not read vendor terms, model disclaimers, or a product's review features as a transfer of responsibility. Nothing in the guidance makes the agent's responsibility conditional on the tool.

Practice takeaway. When AI-assisted advice is wrong, the registered migration agent who gave it is responsible. OMARA's wording is deliberately source-blind — the agent “remains responsible for that information, regardless of the source” — and the statutory chain points the same way: assistance is provided by a person, only certain people may lawfully provide it, and there is no exemption for information generated through AI. The model is not the regulated party and the vendor does not take the error. The path from a wrong answer therefore runs through the agent's own conduct: a complaint to OMARA, an investigation where it has reasonable suspicions of a Code breach, and disciplinary action where appropriate.