What Is AI in Tax? Take the Autopilot Test!

The presenter asked us a question that should have been simple.

"If a 150-million-euro airplane flies itself, takes off by itself, and lands by itself, is that AI?"

The room split. Most people said no. A few said yes. One person gave the answer that turned out to be correct.

"It depends on when you ask."

If you cannot tell a clear story about what AI is, you cannot govern it, buy it, or defend it in court. So let us get the definition right.

Quick answer

AI in tax administration is any system that takes information in, processes it using rules or learned patterns, and produces outputs that would otherwise need human judgment.

By older definitions, even a rule-based audit scorer counts as AI. By modern definitions, the system usually has to learn from data on its own. Both definitions matter in tax work, and the difference shapes how you govern each kind.

The Autopilot Question

Back to the airplane.

The autopilot is sophisticated. It uses sensors to maintain altitude, heading, and speed. It can take off and land on its own. It corrects for wind and turbulence. It is one of the most reliable pieces of software in any industry.

But it does the same thing every time. Same inputs, same outputs. It does not improve with experience. A data scientist would call this a deterministic, rule-based system. That is intentional. In aviation, you want predictability above everything else.

By the old definition of AI, the autopilot qualifies. By the modern definition, it does not.

That gap is the heart of every confused AI conversation in a tax administration.

The Turing Era: Old AI

The original definition of AI came from Alan Turing's 1950 paper in the journal Mind. He described a machine that could imitate human behavior well enough that a person could not tell it apart from another human. If it could do that, he argued, you could reasonably say it had intelligence.

For decades, that was the standard. Anything that imitated human judgment counted as AI. Rule-based systems, expert systems, simple classifiers, all of them qualified.

This is also why "The Imitation Game," the 2014 film about Turing, picked that title. Imitation was the test.

If you brought a 1985 expert system into a 2026 procurement meeting, no one would call it AI. The bar has moved.

The Modern Definition: Can It Learn?

Today the most common test is different. Can the machine learn on its own?

That single question separates classic automation from what most people now mean by AI.

A modern definition usually includes three things:

1. The system takes in information from sensors, data feeds, or user input.
2. It updates its own internal patterns or weights based on what it sees.
3. It produces outputs that change as the inputs change in ways the original developers did not hardcode.

The EU AI Act, adopted in 2024, uses a definition very close to this. So does the OECD's recommendation on artificial intelligence, first published in 2019 and updated in 2023.

This is the definition you should use when writing your administration's AI policy. It is the one regulators are using.

Why This Matters for Tax

The definition is not academic. It has real consequences in three areas.

Procurement. Vendors love to label everything "AI-powered." Once you know the modern test, you can ask the right question. "Does this system learn from our data, or does it apply fixed rules?" Both can be valuable. They are not the same product.

Compliance and audit. If your administration uses a learning system to select audit cases, you owe the taxpayer an explanation. Rule-based systems are easier to defend in court. Learning systems are harder. The Netherlands childcare benefits scandal, which I will cover in detail in post 3, is the most cited example of what happens when an administration cannot explain its own algorithm.

Policy. Your AI policy has to cover both kinds. Rule-based systems need version control, change logs, and a clear owner. Learning systems need all of that plus training data review, bias testing, and ongoing performance monitoring.

The History Most People Miss

Here is the part of the workshop that surprised the room. AI in tax is not new. It is not even close to new.

In most advanced economies, IT systems in tax administrations started arriving in the late 1970s, when punch cards were still common. Core process automation came in the 1980s and 1990s. Self-service portals arrived in the 2000s. Data warehouses and data lakes followed.

AI followed the same arc, on a parallel track:

• 1970s and 1980s: First expert systems and basic classifiers
• 1990s: Neural network research papers in customs
• 2000s: Predictive analytics era. SAS, SPSS, R, and RStudio became common in larger administrations.
• 2010s: Machine learning algorithms went mainstream. Random forests, decision trees, regression trees, and k-nearest neighbors became standard.
• 2020s: Generative AI arrives and rewrites the rules again.

Each of these waves added something. None of them replaced what came before.

The IRS Classifier That Started in 1970

The cleanest example of AI in tax administration is also one of the oldest.

The US Internal Revenue Service has been using a system called DIF, the Discriminant Index Function, since around 1970. DIF is a classifier. It looks at the patterns in tax returns and assigns each one a score that estimates the likelihood of non-compliance. The IRS uses those scores to decide which returns to audit.

By the older definition of AI, DIF clearly qualifies. By the modern definition, it depends on the version. The original DIF was statistical and updated infrequently. Newer iterations behave more like a learning system.

Either way, the point is the same. A national tax administration has been using algorithmic case selection for over fifty years. The conversation about AI in tax did not start with ChatGPT.

Bottom-Up Versus Top-Down

One last distinction will save you time in every future AI conversation.

Machine learning is bottom-up. You start with your data. You let the algorithm find patterns in it. The model is shaped by what you feed it. Data quality matters enormously. So does your domain expertise, since you need to know if the patterns make sense.

Generative AI is top-down. Someone else has already trained a giant model on most of the public internet. You bring it into your environment and apply it. Your data is not what trained the model. Your data is what the model gets pointed at, after the fact.

That distinction has real budget implications. ML projects fail when your data is dirty. Generative AI projects fail when your prompts are vague and your guardrails are weak. They are not the same kind of project.

What This Means for Your Administration

If you take three things from this post, take these.

1. Use the modern definition for your AI policy. "Systems that learn from data" is the line regulators are drawing. Anchor your policy there.
2. Audit your existing systems against both definitions. You probably have older AI in production already. Inventory it before someone else does.
3. Separate ML projects from generative AI projects in your planning. They look similar from the outside. They are different on the inside.

The autopilot question is not a trick. It is a useful test. Walk into your next vendor meeting and ask it. The answer you get will tell you how seriously to take the rest of the pitch.

---

Frequently Asked Questions

What is AI in tax administration?

AI in tax administration is the use of systems that take information in, process it, and produce outputs that would otherwise need human judgment. Modern definitions usually require the system to learn from data on its own. Older definitions include rule-based expert systems.

What is the difference between AI and machine learning?

Machine learning is a subset of AI. It refers specifically to systems that learn patterns from data without being explicitly programmed. AI is the broader category and includes rule-based systems, expert systems, machine learning, and generative AI.

Has the IRS always used AI?

The IRS has used algorithmic case selection through its Discriminant Index Function (DIF) since around 1970. By the standards of that era, DIF was AI. By modern standards, the original version was a statistical classifier, while newer iterations behave more like learning systems.

Is the EU AI Act's definition of AI useful for tax administrations?

Yes. The EU AI Act, adopted in 2024, defines an AI system in a way that aligns closely with the OECD definition. Both emphasize systems that learn or infer from inputs. Most tax administrations writing AI policies today are anchoring on one of these two definitions.

Is generative AI the same as machine learning?

No. Machine learning is bottom-up, meaning the model is built from your data. Generative AI is top-down, meaning you use a pre-trained foundation model and apply it to your work. Both are forms of AI, but the projects, risks, and governance needs are different.

What is the simplest AI definition for a tax administration policy?

A practical working definition is "any system that takes information in, processes it using rules or learned patterns, and produces outputs that would otherwise require human judgment." This covers both rule-based and learning systems and is broad enough to govern.

---

References

1. Turing, A.M. (1950). Computing Machinery and Intelligence. Mind, 49 (236), pp. 433-460. The foundational paper that introduced what later became known as the Turing Test.
2. International Monetary Fund. International Survey on Revenue Administration (ISORA). The primary global source for comparative data on tax administration practices, including AI adoption.
3. US Internal Revenue Service. Discriminant Index Function (DIF). Referenced in the IRS Internal Revenue Manual and in successive reports from the Treasury Inspector General for Tax Administration. DIF has been used for audit case selection since approximately 1970.
4. European Union. Regulation (EU) 2024/1689 (the EU AI Act). Adopted in 2024. Article 3 contains the formal definition of an "AI system."
5. OECD (2019, updated 2023). Recommendation of the Council on Artificial Intelligence. The most widely cited international definition of an AI system outside the EU AI Act.
6. Russell, S. and Norvig, P. (2020). Artificial Intelligence: A Modern Approach. 4th edition. Pearson. The standard university textbook on AI, useful for grounding internal training material.
7. International Monetary Fund. Technical Notes on AI in Revenue Administration. Available through the IMF eLibrary. Practical guidance for revenue administrations adopting AI.
8. The Imitation Game (2014). Directed by Morten Tyldum. Dramatization of Alan Turing's life and his work on the Enigma cipher during the Second World War.

---

Suggested internal link ideas

These point forward and back across the series.

• Where tax administrations actually stand with AI (post 1 in this series)
• When AI goes wrong: the Netherlands childcare benefits scandal explained (post 3, coming soon)
• Generative AI for tax administrations: what is actually different this time (post 4)
• AI agents in tax and customs: the next shift (post 5)
• How to risk-assess an AI use case in a tax administration (post 6)

Suggested external authority sources

For readers who want to go deeper into the definitions.

• IMF eLibrary, especially technical notes on AI in revenue administration
• OECD AI Policy Observatory, including the AI Principles
• EU AI Act portal maintained by the European Commission
• IRS Internal Revenue Manual sections on case selection
• David Hadwen's public AI use cases database for tax administrations
• The Alan Turing Institute, particularly its work on AI governance

If this clarified the definition for you, the next post in the series is where the definition gets tested in the real world. I unpack the Netherlands childcare benefits scandal, the HMRC chatbot story, and the question every tax administration has to answer in court. Why did the algorithm make that decision?

Read it next, or subscribe so you do not miss the rest of the series.