Submission 79
Tsutomu Yahagi
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
A Japan-based tax and labour consultant argues that AI should be assessed on its ability to decline answers it has no authority to give, and that the verification burden is the main SME barrier. Proposes practitioner-written limits, refusal logs and clarity that AI output is not professional judgement. States they hold patent applications, five in Australia, on constraining AI output.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
The submitter argues that the main barrier to SME adoption is the verification burden, not cost, skills or technology capability.
The Committee asks what is stopping small and medium businesses from adopting AI. Seen from inside a practice serving 175 of them, the answer is not the usual one. It is not cost. It is not skills. It is not that the technology is insufficient. It is the verification burden.
They argue that domain-specific training improves accuracy but not authority, and that such models also need a specification of what they must not produce.
Accuracy and authority are different properties. Training improves the first and leaves the second untouched. What a professional-domain model additionally requires is a specification of what it must not produce.
They argue that professional registration and liability frameworks do not currently distinguish a judgement a practitioner formed from a model output the practitioner passed on.
Regulated professions rest on an identifiable person being answerable. A tax agent signs. A lawyer advises. A doctor diagnoses. Both the registration regime and the liability regime assume that a human judgement occurred and can be located. Nothing in that framework currently distinguishes between a judgement a practitioner formed and a model output the practitioner passed on.
They argue that, in professional contexts, a record of what a system declined to produce is more useful than a record of what it produced.
At present systems are required to record what they produced. In professional contexts the more useful record is what the system did not produce, and on what basis. A refusal log is small, auditable, and bears directly on whether the constraint was operating.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Make the ability to decline an answer a criterion when AI is assessed.
Make the ability to decline an answer a criterion when AI is assessed.
Where public support is given, have practitioners carrying liability write specifications of what must not be produced, rather than only checking outputs.
Where public support is given, put practitioners on the side that writes the specification. Asking them to check the output is not enough. What must not be produced can only be written by someone who carries the liability.
Consider requiring systems used to prepare professionally signed documents to record what they declined to produce.
Consider requiring systems used to prepare documents a professional signs to keep a record of what they declined to produce.
Consider clarifying that "the AI said so" does not constitute the exercise of professional judgement; the stated purpose is not to create new liability.
Consider clarifying that "the AI said so" does not constitute the exercise of professional judgement. The purpose is not to create new liability. It is to stop "I formed a judgement" and "I passed on the model's answer" from being treated as the same thing.
Address the verification burden rather than model capability if SME adoption is the goal.
If adoption by small and medium business is the goal, address the verification burden rather than model capability.