Artificial Intelligence in 2023: Impact and Ethics

2023 was the year artificial intelligence stopped being a research topic and became a line item. Generative models moved from demonstrations into ordinary software: the spreadsheet, the email client, the bookkeeping package, the customer service queue.

Three years on, the interesting question is no longer whether the technology works. It is what a Canadian business actually takes on when it puts AI into a process that produces its books, its client records or its tax filings. That is a narrower question than the ethics debate, and it is the one with money attached.

What 2023 actually established, and what it did not

It established that a general-purpose model can produce usable first drafts of routine work: correspondence, summaries, classification, code. It established that the marginal cost of that work fell sharply.

It did not establish that the output is reliable without review, and it did not change a single obligation in the Income Tax Act. The claims made in 2023 about diagnosis, manufacturing and creative work were mostly forecasts. The obligations described below were law before the forecasts and remain law after them.

Your records are still your records

This is the point most owners miss when they adopt an AI-assisted bookkeeping or document tool.

The CRA requires you to keep books and records at your place of business or your residence in Canada, unless it gives you written permission to keep them elsewhere. Records held on a server outside Canada and merely accessed electronically from within Canada are not records in Canada. Most AI tools are cloud services, and most cloud services store data wherever the vendor chooses.

The CRA’s electronic record keeping circular adds a second requirement that AI tooling makes easy to fail: the records must remain in an electronically readable and usable format, with enough detail to support the returns you filed. A tool that summarises a receipt and discards the image has destroyed the record and kept the summary.

The practical rule: whatever the tool does, you keep the underlying source document, in Canada, for six years from the end of the last tax year it relates to. See digital record keeping for the CRA for how that works in practice, and bookkeeping automation for where automation genuinely helps.

How AI spending is treated on the return

Owners routinely assume that anything involving AI is either an ordinary expense or a research credit. It is usually the first, occasionally the second, and sometimes neither.

The first and third rows are the same decision seen from two sides, and the choice between them, hosted service against a model you download and run yourself, is worked through in DeepSeek against ChatGPT.

The classes and rates for depreciable property are set by regulation and change, so check the current capital cost allowance classes for the year you are filing rather than working from a figure you read somewhere.

Where AI work becomes an SR&ED claim, and where it does not

This is the most misunderstood interaction in the whole subject, and it is expensive to get wrong.

SR&ED requires two things at once. The work must be undertaken for the advancement of scientific knowledge or a technological advancement, and it must be a systematic investigation carried out by experiment or analysis in a field of science or technology. The CRA’s eligibility guidelines turn on whether there was a technological uncertainty at the start: something that could not be resolved from the existing knowledge base by a competent practitioner.

Adopting a model is not that. Prompting a model is not that. Integrating an API into your product, however difficult the deadline, is not that. What can qualify is work where you did not know at the outset whether the result was achievable, you formulated a hypothesis, and you tested it and recorded what happened.

The documentation is the claim. Contemporaneous notes of the uncertainty, the hypotheses, the experiments and the results are what survives a review. A reconstructed narrative written after the fiscal year end rarely does. The same pattern applies here as in SR&ED claims generally.

The ethical questions that carry financial consequences

The 2023 conversation about AI ethics was largely abstract. Two parts of it are not.

Personal information. If your AI tool processes client names, contact details, financial information or health information in the course of commercial activity, PIPEDA applies to that processing, and it applies whether or not you knew where the vendor sends the data. Consent, purpose limitation and safeguarding obligations do not pause because a third-party model is in the pipeline. Read the vendor’s data-processing terms before the tool touches a client file, not after.

Accountability for output. A model that mis-classifies an expense, invents a citation or drops a transaction produces an error attributable to you. The CRA assesses the taxpayer, not the software. Gross negligence penalties turn on your conduct, and “the tool did it” is not a defence that has ever worked for a bookkeeping package. Review AI output on anything that ends up on a return.

Bias matters most where a model influences a decision about a person: lending, hiring, pricing. If you are not making those decisions, the bias question is real but not yours. If you are, it is a legal exposure and not just an ethical one.

A short decision for a business owner

Is the AI touching data that ends up in your books or on a return?
├─ No  → treat it as any other software purchase
└─ Yes → 1. Where is the underlying record stored?
         2. Is the source document retained, not just the summary?
         3. Who reviews the output before it is relied on?
         4. Does the vendor contract cover personal information?

Four questions, and they take an afternoon. They are worth more than any projection about what AI will do to the economy, because they are the ones a CRA reviewer will actually ask about.

Where AI is genuinely changing professional work is described in AI in accounting in Ottawa, and the security side of adopting new tooling is covered in cyber security for small business finances.

What to do this quarter

Inventory the AI tools already in your business, including the ones that arrived inside software you already had. For each, answer where the data lives, what source documents are retained, and who checks the output. Then decide whether any of the development work you paid for was genuinely uncertain at the outset, because that determination has to be made while the people who did the work can still remember it.

If you have adopted AI tooling in a process that feeds your books and you are not sure whether your records still satisfy the CRA, that is worth a specific review rather than a general reassurance.