Garbage In, Garbage Out, and the Layers of Interpretation

Garbage in, garbage out is an old data principle with renewed importance in AI-enabled work. AI systems can generate polished answers from weak instructions, incomplete evidence or poor-quality data. The risk is that the output looks credible enough to be accepted, even when it is generic, inaccurate or unsuitable for the situation. Carmichael (2026) argues that good usability can make AI feel easier than it really is: using AI well requires engagement before and after the tool generates a response.

The AI Sandwich model provides a practical way to think about this. The first human layer is the brief. A strong brief explains the purpose, the decision to be supported, the audience, the required format, the data available, the boundaries, the risks and the standard of evidence expected. This might mean telling an AI tool whether the task is exploratory, advisory, analytical or decision-supporting. A request such as ‘write a performance report’ is too vague. A better brief explains the performance question, the measures being used, the time period, the stakeholders affected and the judgement criteria. The quality of the input shapes the quality of the AI response.

The middle layer is the AI filling. This is where the tool can be useful: summarising information, comparing options, drafting first versions, identifying possible risks, structuring questions or offering alternative explanations. However, AI output is not the same as organisational truth. Generative AI predicts likely responses from patterns in data; it can produce confident but false information, known as hallucination (GOV.UK, 2025). It may also reproduce bias, omit context, misunderstand intent or over-simplify complex workplace issues. NIST (2023) describes AI risk management as a way to improve trustworthiness across design, development, use and evaluation, which means practitioners should treat AI output as evidence to examine, not a conclusion to copy.

The second human layer is curation. This is where layers of interpretation become essential. Practitioners should ask: What does this output assume? What evidence supports it? What has been left out? Who could be affected? Is the tone appropriate? Does it reflect the organisation’s policy, values and operating context? Would the same recommendation still make sense if the data were incomplete or if a vulnerable group were affected? This interpretation stage protects quality because it reconnects the output to professional judgement and workplace reality.

Data quality is central to the whole sandwich. DAMA International (2017) describes data management as a discipline concerned with treating data as an asset, including quality, governance, security and meaning. In AI work, practitioners need to consider whether data is accurate, complete, current, consistent, relevant and lawful to use. For example, an AI tool asked to analyse workforce absence may produce misleading themes if absence reasons have been recorded inconsistently, if part-time staff are not comparable with full-time staff, or if sensitive data is being used without clear controls. The ICO (2023) links AI accuracy and fairness to the wider data protection lifecycle, including accountability, transparency and the risk of unfair outcomes.

Responsible interpretation also means knowing when AI should not be used, or when its role should be limited. Daugherty and Wilson (2018) argue that value is created when people and intelligent machines work together, with people retaining roles that require judgement, empathy, ethics and context. In practice, this means AI can support the thinking, but the practitioner remains accountable for the final message, recommendation or decision. GOV.UK (2025) is clear that AI systems need testing, monitoring, assurance and meaningful human control at the right stages. The sandwich therefore becomes a governance habit: define carefully, generate critically, curate responsibly and document what has changed.

The strongest use of AI is conversational rather than transactional. A user may refine the prompt, challenge the output, ask for missing assumptions, request counterarguments, compare risks and then check the final version against verified sources and organisational knowledge. This keeps human expertise visible. It also prevents a common failure: using AI to produce quick content while quietly outsourcing the thinking. Garbage in, garbage out reminds practitioners that input quality matters. Layers of interpretation remind them that output quality is earned through questioning, checking and ownership.