# AI Transparency: Can We Know Whether We Are Interacting With a Human or a Machine?

Author: Norbert Biedrzycki
Published: 2025-11-15
Language: en
URL: https://norbertbiedrzycki.pl/en/ai-transparency-can-we-know-whether-we-are-interacting-with-a-human/
Summary: In October, Deloitte Australia agreed to partially refund the Australian government for a report worth 440,000 Australian dollars, reviewing the IT system behind automated penalties in the welfare…

In October, Deloitte Australia agreed to partially refund the Australian government for a report worth 440,000 Australian dollars, reviewing the IT system behind automated penalties in the welfare system. An academic had found that the report contained references to research papers that did not exist and a fabricated quote from a federal court judgment. A revised version disclosed that a generative AI tool had been used in its preparation. Deloitte stood by the report's findings and recommendations.

The irony is hard to miss. A report about the risks of automated decision making in welfare, an area where Australia has painful experience after the Robodebt scandal, was itself undermined by undisclosed automation.

## Two transparency problems

The case illustrates two different transparency questions that are often mixed together.

**Disclosure of use.** Did the people relying on a piece of work, a decision or an interaction know that AI was involved?

**Disclosure of nature.** Is a particular piece of content, an image, a voice or a message, authentic or synthetic?

Both are becoming urgent. Customers increasingly interact with AI systems without knowing it. Professional work products are increasingly cowritten by models. Synthetic images and voices are now indistinguishable from real ones for most people.

## Where regulation is heading

The EU AI Act addresses both. From August 2026, its transparency obligations will require that people are informed when they interact with an AI system, unless it is obvious, and that synthetic audio, image, video and text content is marked in a machine readable way and detectable as artificially generated. Deepfakes must be disclosed. The Commission is working on a code of practice for marking and labelling AI generated content.

Technical standards are developing in parallel. The C2PA content credentials standard attaches verifiable information about the origin and editing history of a file. Google's SynthID embeds watermarks in AI generated content. None of these solutions is complete, and watermarks can be removed. But together they are beginning to create an infrastructure of provenance.

## Beyond labels

Transparency is often treated as a labelling exercise: add a notice saying "generated with AI" and the obligation is met. That misses the purpose.

The point of transparency is to allow people to make informed decisions. A customer should know whether the advice they are receiving comes from a system that may be wrong in specific ways. A client paying for a professional report should know how it was produced and what was checked. An employee receiving a message from leadership should be able to trust its source.

For professional services firms, the Deloitte case raises a sharper question. When clients pay for expertise, they are paying for judgment and verification, not only for text. Using AI is not the problem. Using it without disclosure and without rigorous checking undermines the very thing the client is buying.

For organizations, a workable rule is to disclose AI where it would change how a reasonable person interprets or relies on the output. That covers customer interactions, professional deliverables and anything presented as evidence. It does not need to cover every spellcheck.

> *Leadership question: Should companies be required to disclose every meaningful use of AI, and what level of transparency is necessary to maintain public trust?*
