2.2 Fabrications, Reliability, and Bias

2.2 Fabrications, Reliability, and Bias

Quick Answer: Generative models predict fluent next tokens; they do not look up truth unless you ground them and review them. Fabrications (hallucinations) are confident, unsupported statements. Reliability is consistent, safe behavior across people and edge cases. Bias is systematic unfairness in who is helped or harmed. Leaders do not ship ungrounded answers to customers, and they add extra controls in high-stakes domains.

Microsoft's AB-731 skills list names three challenges of generative AI solutions: fabrications, reliability, and bias. Microsoft's responsible AI materials warn that without careful design, generative models can produce incorrect or even harmful content. Your job as a transformation leader is not to become a research scientist. It is to refuse "the demo sounded sure" as a production standard.

Fabrications (hallucinations)

A fabrication, often called a hallucination in industry language, is output that is fluent and plausible but not supported by facts, by your documents, or by the user's actual request. It can invent a policy clause, a product SKU, a case citation, a refund rule, or a comforting statistic. The model is doing what it was trained to do: continue text. Microsoft's prompt-engineering guidance reminds practitioners that the model produces what is most likely, not a separate "truth" program.

Fabrications are especially dangerous when:

• The user cannot easily check the claim (a customer on a public site, a new hire, a non-expert).
• The tone is confident (people over-rely; Microsoft has written about overreliance as a leadership concern alongside hallucinations).
• The domain has legal, safety, or financial consequences.

Grounding reduces—not eliminates—this risk. Microsoft's Foundry RAG documentation is blunt: if retrieval returns weak passages, the model can still be incomplete or inaccurate despite grounding, and models can still fabricate even with retrieved content. That is why citations, "I do not know" behavior, and human review remain part of the design.

Reliability

Reliability (paired with safety in Microsoft's responsible AI principles) means the system performs consistently for different people, languages, devices, and messy real inputs—not only the happy-path script used in an executive demo. Unreliable generative systems show up as:

• Monday's answer contradicts Friday's answer on the same policy PDF
• English users get precise citations while another language gets vague advice
• A Copilot in Word drafts well, but the same prompt in a custom agent omits disclaimers
• Peak load causes timeouts, truncated answers, or a silent fallback to an ungrounded model

Reliability is an operations problem as much as a model problem: evaluation sets, monitoring for groundedness, incident response when a bad answer ships, and a rollback path. Microsoft's reliability-and-safety principle asks how a system functions in conditions it was not originally intended for. That is the question an operations vice president should put on the steering-committee slide.

Bias, without replacing the later responsible AI chapter

Bias here means systematic difference in quality or harm across groups of people—not a single rude sentence. Examples a leader will actually see:

• An HR screening assistant that writes stronger summaries for resumes that resemble the incumbent workforce
• A customer bot that is less accurate for dialect, accent-transcribed text, or non-English names
• Marketing copy or images that default to stereotypes
• A ranking or "priority" suggestion that correlates with protected characteristics even when nobody typed those words

Microsoft's fairness principle states that AI systems should treat people fairly in how they allocate opportunities, resources, and information. Copilot Studio's responsible AI guidance tells makers to use diverse, representative data, watch for biased outputs, and use human-in-the-loop review. You will go deeper on Microsoft's full responsible AI standard—fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability—in a later chapter. In this chapter, treat fairness as a go/no-go risk on any generative workflow that ranks people, allocates benefits, or speaks in the company's name.

Two mitigations to insist on (preview of grounding and RAG)

1. Grounding

Grounding means giving the model authorized source material to draw from—policy pages, product specs, knowledge bases—and instructing it to stay there. Microsoft recommends grounding whenever the use case needs up-to-date, reliable information rather than pure creativity, and notes that the closer the source is to the desired answer, the less room there is for error.

Retrieval-augmented generation (RAG) is the common architecture: retrieve relevant chunks from an index (often Azure AI Search in Microsoft Foundry), augment the prompt with those chunks, and generate an answer that can cite them. RAG is the subject of the next chapter. What you need now is the leadership rule: no grounding plan means no customer-facing factual assistant.

2. Human-in-the-loop

Human-in-the-loop (HITL) means a person reviews, edits, or approves output before it causes an irreversible action. Intensity should match harm.

Transparency belongs next to HITL: people should know they are using generative AI, which Microsoft calls out in Copilot Studio responsible-AI guidance.

Leader responses

A transformation leader's default answers:

• Do not ship ungrounded answers to customers. Fluency is not a warranty. If the knowledge is not in an approved index (or an equivalent grounded connection such as Microsoft Graph content the user is allowed to see), the product should refuse, escalate, or hand off to a person.
• High-stakes domains need extra controls: mandatory citations, specialist reviewers, evaluation before launch, abuse and content-safety filters, access control at retrieval time (RAG can leak whatever you indexed), and a kill switch.
• Security and legal sit at the table: prompt injection via documents, over-permissioned data, and "the model said it" are organizational incidents, not amusing edge cases.
• Fairness is in scope even before the RAI chapter: if HR, credit, or customer-priority workflows treat groups differently, pause the rollout and bring the later governance work forward rather than arguing that "it is only a chatbot."

If a vendor promises "zero hallucinations," treat that as a sales claim. Microsoft's public materials discuss managing hallucinations and overreliance—not wishing them away. Your independent OpenExamPrep study point is simple: controls first, then scale. Scaling an ungrounded answer is how a fabrication becomes a brand event.

A general counsel reviewing a public FAQ bot should ask three questions in the same meeting: What is the source of truth? Who can stop a bad answer? Which groups of customers will this fail first? Those three questions cover fabrications, reliability, and bias without waiting for a longer responsible AI workshop.