AI literacy is knowing when a draft stops being your thinking

In June 2026, BCG published a survey of 70 C-suite and senior executives on how AI is changing their teams. Half said they were already seeing de-skilling. More than 60% expected it to become a material threat within three to five years. Almost 90% pointed to the same starting point: people accepting AI output without stress-testing it. One leader described the pattern as teams treating a generated draft as 'good enough' and skipping the critical thinking step.

That is a more useful definition of AI literacy than the one that circulates on LinkedIn. Prompt tricks, model comparisons and tool stacks are the easy part. The harder skill is catching the exact moment a draft stops being your thinking and starts being someone else's. BCG's leaders are describing people who have lost that reflex, or never built it.

Where the judgement actually goes

The mechanism is specific. When you draft without AI, every sentence forces a small decision: what you actually believe, what you are willing to defend, what you do not yet know. Those decisions are how judgement gets built. BCG's respondents named the same loss in operational terms. 53% cited slower development of junior talent, because AI now absorbs the research, drafting and problem-decomposition work through which that judgement was traditionally formed. More than half cited reduced ownership and accountability, with employees deflecting responsibility onto the model once they stopped interrogating its outputs. 49% cited less diversity of thinking and 43% fewer constructive debates, as teams using the same tools trained on the same corpora converged on the same formats.

For a professional whose reputation depends on published work, this matters at the sentence level. A post, a memo, a client note or an internal strategy document is a public commitment to a view. If the view in it is the model's statistical average of similar documents, you have published a position you did not actually form. Readers who know the territory can usually tell. The sentences are fluent and the structure is tidy, but the specific judgement that would make the piece worth citing is missing. That is what people on LinkedIn mean when they complain about 'slop', and it is why reach for legitimate voices is being affected. The feed is being trained to distrust the shape.

A test to run before you publish

The practical test is simple. Read the draft and underline every claim that goes beyond description. For each one, ask: did I decide this, or did the model? If you decided it, the sentence stays and probably gets sharper. If the model decided it, you have three honest options. Rewrite it in your own words, having thought about it. Replace it with a narrower claim you can defend. Cut it. There is no fourth option in which you keep the sentence because it sounds right.

BCG frames this as a system-design problem rather than a talent problem, which is correct for organisations. They recommend AI-off zones for certain tasks, prompts that force the model to argue against itself ('give me a solution that does not work'), reflective nudges inside tools, and keeping final judgement with a human. The individual version of the same discipline is a short policy you apply to your own output. Decide which categories of writing you will not draft with AI. Decide which categories you will draft with AI but rewrite in full. Decide which you will publish with light edits. Write it down in advance, before the deadline pressure arrives. The point is not purity. The point is that the decision has already been made, when you were thinking clearly.

On the evidence, and the cost of getting it wrong

A note on the source. BCG's sample is 70 executives and the findings are self-reported perceptions rather than measured skill tests. The piece also covers AI use at work generally, not professional publishing specifically. The extension to posts, memos and client notes is my reading, not BCG's claim. The leader observations still line up with what is visible on any professional feed and in most inboxes: more volume, more fluency, less distinct thinking.

The reputational cost of getting this wrong is quiet and cumulative. No single AI-assisted post ruins a reputation. A year of them, published without the step where you decide what you believe, produces a public body of work that reads like everyone else's. The audience you were trying to reach stops reading before they stop following. By the time the reach numbers show it, the position has already been lost.

AI literacy, in the version worth having, is a reading skill applied to your own drafts before anyone else sees them. It is not about using less AI. It is about catching, sentence by sentence, when the thinking stopped being yours. The leaders in BCG's survey are watching that reflex erode across whole organisations. The same reflex is what keeps an individual professional's published work worth reading.

Source: BCG, 'When Everyone Uses AI, Companies Risk Losing Critical Skills' (17 June 2026).

Source: https://www.bcg.com/publications/2026/when-everyone-uses-ai-companies-risk-critical-skills