search for news about generative ai and jobs these days and you'll probably get headlines that feel like they belong in the necrology section of a newspaper, announcing which job just died (or, if it's a viral tweet, linkedin post or youtube video it's more lilely to say the job is "cooked", "finished", "done", "extinct", "toast", "over", "bust"), and which one is "next".

i am clearly not safe from the temptation of a catchy headline either of course (just read up).

these headlines make it sound like ai coming for our jobs is inevitable, and imminent.

this is not new: an exciting technology appears, and we conclude it will wipe out jobs almost overnight.

ai is just the latest chapter in that story.

the data

mckinsey's new report titled: "agents, robots, and us: skill partnership in the age of ai" states that today’s technologies (ai agents, robots and workflow redesign) could in theory automate 57 percent of current work hours in the u.s. (44 of that 57% coming from ai agents).

it is important to note that it does not say 57% of jobs. and it specifies "in theory" because those numbers assume perfect technical conditions which almost never match how work happens in real life.

so the technical potential of an innovation and its real adoption in the workplace are two different things. electricity took more than three decades to become standard in factories even though it was clearly better than steam.

robotics, cloud, and machine learning followed a similar paths. organizations take time to redesign systems, train people, and adjust responsibilities.

what this means for jobs

ai and automation will likely reshape certain jobs, and some roles will likely decline, as new ones emerge but the net impact on labor is hard to predict.

the oecd estimates that about 14 percent of jobs across member countries are highly automatable, concentrated in routine administrative, clerical, and predictable production work.

but yale budget lab concludes that “measures of exposure, automation, and augmentation show no sign of being related to changes in employment or unemployment” and that “better data is needed to fully understand the impact of ai on the labor market”.

then goldman sachs, states that “ai adoption is expected to have only a modest and relatively temporary impact on employment levels”.

yet forbes recently published an article titled "ai is not killing entry level jobs" based on research that concludes looking at the the declining youth employment assumed to be caused by ai that the: "junior hiring crash started in 2022, before many ceos knew what prompt engineering was."

so there is really no consensus on how ai will or will not impact employment rates

that is probably because this will be driven by many factors beyond the technology’s capabilities such.

• psychology (loss aversion, confirming or anchoring biases),

• policy (labor regulation, education and retraining systems, social safety nets),

• economics (wage dynamics, productivity distribution, capital investment),

• and ethical and moral choices (fairness, responsibility, and how societies value human work).

• etc.

  any definitive statements on how ai will impact job numbers should be taken with a grain of salt.

the immediate opportunity is in task removal

what the data does show is that ai in its current and foreseeable state can affect what people do, rather than whether they are needed.

mckinsey’s analysis shows that while 57% of work hours contain activities that could be automated, 43% still require social and emotional capabilities that current systems cannot perform. these activities rely on sensing client uncertainty, navigating conflict, reading context, and building trust.

how generative ai works

so why can’t ai take over fully and exclusively these relational skills when modern models can identify tone and mood from our inputs?

sentiment classifiers can flag frustration. chatbots can respond with language such as i understand why that feels difficult. this can give the impression of understanding.

but the underlying mechanics are different from what people generally implicitely assume when they say understanding.

this piece by mit highligths that large language models operate through next token prediction. they select the most statistically likely token based on patterns in their training data. a token is the fundamental unit of data used to process and generate text.

ai models can't directly "read" text like humans, they break down language into smaller chunks called tokens, which are then converted into numerical values for mathematical processing.

llms do not interpret meaning, researchers explain, they work with tokens as numerical vectors, not grounded semantic concepts. they do not know what frustration is; they only recognize the patterns of language associated with it.

linguists describe this gap as a difference between sign and meaning. in simple terms, humans link words to lived experience, context, and shared understanding. llms do not attach meaning. they generate statistically coherent sequences based on correlations in large datasets.

this gap also explains "hallucinations" in generative ai.

this nature article explains that, contrary to belief, hallucinations (when chatgpt or your favorite conversational ai tells you something that is not true) aren't bugs or malfunctions, they are inherent to how llms are designed on matching patterns and predicting the next word from their training data.

so, llms are trained on vast text data, not facts. they generate plausible-sounding text by filling gaps with statistically likely words, essentially fabricating details that fit the context but aren't necessarily true

the most statistically coherent sequence does not necessarily lead to the right decision.

this limitation matters in practical settings. ai can surface signals. it can predict likely responses. but it cannot take responsibility, interpret intent, or sense nuance through experience. these tasks still require people.

letting an unconstrained llm run a business is a risky proposition.

but…