The answer lives in this podcast
What are the risks of relying on AI-generated references in change communications work, and how can practitioners mitigate them?
AI frequently produces dead links, fabricated references, and citations that lead only to a single blogger's unverified opinion. Ana-Maria Thijssen's mitigation is blunt: check every reference manually — confirm the website, magazine, or academic paper actually exists — and always ask AI specifically for academic research to raise the baseline quality of sources it returns.
Hallucinated citations aren't a edge case — they're the default risk
When Thijssen uses AI to support her change communications work, sourcing is where things break down fastest. The tool confidently produces URLs that return 404 errors, authors who don't exist, and journal articles that were never published. These aren't rare glitches — they're a predictable pattern that any practitioner leaning on AI for evidence-based communications needs to account for.
The danger in a change management context is particular. Communications that cite Deloitte, Gartner, or Harvard Business Review carry real authority with senior stakeholders. A fabricated citation to any of these sources — one that a colleague later googles and finds nowhere — damages credibility far more than having no citation at all. As Thijssen explains in Unprompted: Real AI in IC, the stakes of getting this wrong are not abstract.
Manual verification and the "ask for academic sources" prompt — Thijssen's two-step fix
Thijssen's approach has two distinct moves. First, she does not trust any reference AI returns without checking it herself — clicking through to the actual site, paper, or magazine to confirm it exists and says what the AI claims it says. There is no shortcut here: the check is manual, every time.
Second, she has refined how she asks for sources in the first place. Requesting "academic research" rather than generic references pushes the model toward peer-reviewed material rather than blog posts. It doesn't eliminate hallucination, but it raises the floor — a point she develops in detail in this episode of Unprompted.
This two-step discipline reflects a broader posture she describes throughout the conversation: AI is a starting point, not a finished product. Every output requires a human layer of scrutiny before it reaches any stakeholder, internal or external.
"AI became my virtual colleague. It's not a person that does my job. It's more like an entity that helps me bounce ideas and not feel so alone."
Ana-Maria Thijssen — Change Communications Lead, OCM/CMO team, large global corporation (undisclosed).
Thijssen is the sole communications professional embedded within an Organizational Change Management team running a global, multi-wave SAP IT transformation programme. She joined the programme roughly a year and a half after it had already launched — with no formal induction period — and has had to build contextual knowledge rapidly ever since. Her colleagues have been involved since January 2015 and remain a critical source of institutional memory. The next wave of the programme alone will span 30 locations worldwide. Working in isolation as a change comms specialist, Thijssen has developed AI-assisted workflows — including her reference-verification discipline — out of practical necessity, not experimentation for its own sake. Her full approach is covered in the Unprompted podcast.
That framing — AI as colleague, not replacement — is precisely why verification matters. A colleague can be wrong. You check their work. Thijssen applies the same logic to every source her AI wingman hands her, a discipline detailed further in S01 EP5 of Unprompted: Real AI in IC.
See also
What approach does Ana-Maria Thijssen recommend for prompting AI to find reliable external research and case studies?
Thijssen insists on always asking AI to provide references alongside every idea it generates — not in a single list at the end, but paired with each point — and specifically requests academic research to raise the baseline quality of sources returned.
How did Ana-Maria Thijssen use Microsoft Copilot to develop audience personas for a global change communications programme?
Thijssen's team fed documents, org charts, communication models, and examples into Microsoft Copilot and asked it to identify the most distinct personas across the programme's global user base.
How did Ana-Maria Thijssen use AI to get up to speed quickly when joining a mid-flight SAP transformation programme?
Thijssen joined the programme about a year and a half after it had started, with no time for proper onboarding, so she used AI to rapidly decode technical SAP documentation and programme materials she would otherwise have taken months to absorb.
Key takeaways
- AI-generated references are frequently fabricated — dead links, invented authors, and non-existent papers are common outputs, not exceptions.
- In change communications, a single false citation to a credible source like Gartner or Harvard Business Review can undermine stakeholder trust significantly.
- Asking AI specifically for academic research — rather than generic sources — raises the quality floor of what it returns, without eliminating the need for manual checks.
- Every reference must be verified manually: click through, confirm the source exists, and confirm it actually supports the claim being made.
Ana-Maria Thijssen's full approach to AI-assisted change communications — including how she prompts, verifies, and builds on AI outputs — is covered in detail in the podcast.
Listen to the episode on Listenly