Finding a first place to learn
I work in IT consulting in Lisbon, placing consultants with large clients. Of the 70 positions currently open with my clients in Portugal, two are for people with zero to two years of experience. Clients have always mostly wanted people who can move their projects forward quickly, which makes sense to me: if they bring in a consultant, they are paying for skills, while they can choose to train the people they hire in-house.
What has changed is how often junior requests come in. I used to receive them regularly; today they are clearly rarer. I see it in my own work, without a data series that would let me measure the drop or attribute it to AI. And without suitable assignments, we hire very few beginners ourselves.
I could not find any public series tracking junior hiring in Portuguese tech. The Tech Talent Trends 2026 report by Landing.Jobs and Damia, one of the most detailed snapshots of the local market, still offers a reference point: according to their survey, only 12% of tech professionals in Portugal have less than three years of experience, and company demand is concentrated on senior profiles. Professionals choose whether to answer, so this is a self-selected sample, not a census. [2]
In France, researchers from HEC and the Banque de France looked at hiring according to how exposed each occupation is to AI. Between November 2022 and March 2026, the trend in permanent-contract hires of 21 to 29 year-olds fell by 28.3% in the most exposed occupations, against 7.5% in the least exposed. “Exposed” here means an occupation whose tasks could be performed or assisted by AI, according to an index built by the authors: a technical potential, not a measure of how each company actually uses the tools. [1]
In France, youth hiring falls further in AI-exposed occupations
How to read: permanent-contract hires of 21 to 29 year-olds in the least and most AI-exposed occupations. Zero is November 2022.
Takeaway: −7.5% against −28.3%. The gap does not prove AI is the cause.
Banque de France / HEC · October 2026. Change in the hiring trend, not in jobs. Potential exposure; causality not established.
In the United States, a Stanford team tracks the payroll records of several million employees. As of June 2026, employment of 22 to 25 year-olds in the most AI-exposed occupations stands 19% below where it would be had it kept pace with less exposed occupations, with no comparable gap for more experienced workers. The decline comes mainly from fewer hires, not more departures. One detail stopped me: it is concentrated in occupations where AI replaces tasks, while where AI supports the work, employment is stable or rising. The authors present these as descriptive indicators, not a causal effect. [3]
None of these studies can separate the effect of AI from the economic cycle. But they converge on a point that brings me back to learning: part of what we know how to do is acquired in the assignments someone agrees to give us when we are starting out.
What my first jobs taught me
I think back to my first affiliate marketing dashboards and reports, which I had to produce in Excel and VBA because the platform could not do everything yet. I was frankly bad at Excel, which we had barely covered at school, and it was by building those files and fixing my mistakes that I learned to use it.
With AI, I would probably have saved a lot of time. But I know myself: once the report was done, I would most likely have moved on to the next one without trying to understand each formula. That is what interests me about people starting out today: how do you benefit from the tool while still acquiring what lets you understand the work and take it over later?
Two studies, two different stories
One experiment with 52 programmers asks the question directly: they discover Trio, a Python library, half of them with a GPT-4o assistant and half without AI. On the second task, all 26 assisted participants finish within 35 minutes, against 22 in the other group. But on the comprehension test taken afterwards without the assistant, their average score is around 50 out of 100, against 65 for the no-AI group. [4]
In this experiment, finishing more did not mean understanding more
How to read: the squares count people who completed the second task. The bars below show the average score on the comprehension test without AI.
Takeaway: finishing the task and understanding the code are two different outcomes.
Shen & Tamkin · How AI Impacts Skill Formation. Exercise on an unfamiliar library; most participants already code. Understanding measured immediately.
On its own, that result suggests AI holds back learning. Another study says almost the opposite. At a US software company, Erik Brynjolfsson, Danielle Li and Lindsey Raymond followed 5,172 customer support agents who were gradually given an AI assistant. New agents who had it from their first month reached about 2.5 resolved issues per hour after two months; without the assistant, it took them eight to ten months. [5]
The most interesting detail is elsewhere. During outages of the tool, agents who had used it stayed faster than before it arrived, and the effect was clearly stronger for those who had actually tried its suggestions with customers. The authors read this as lasting learning, while acknowledging that they cannot always tell learning apart from following a recommendation. [5]
The two studies do not measure the same thing: immediate understanding after an hour on one side, output over several months on the other. But they raise the same question: what does the person do with the help they receive?
What separates those who learn from those who just deliver
In the Trio experiment, the researchers watched every participant’s screen recording and sorted their use of AI into six patterns. Three amount to delegating: having the assistant write all the code, sliding gradually into full generation, or asking it to fix errors without trying to understand them. The twelve people concerned scored between 24 and 39 out of 100 on the test. The other three amount to questioning: asking only conceptual questions, requesting an explanation with every piece of code, or generating code and then quizzing the assistant about it. The twelve people concerned scored between 65 and 86 out of 100, on a par with the no-AI group. Those who asked only conceptual questions were even among the fastest. [4]
A field experiment with nearly a thousand Turkish high school students in mathematics points the same way, at a larger scale. Some practised with an interface similar to ChatGPT, others with a tutor set up to give hints rather than answers. During practice, grades rose by 48% with the open AI and by 127% with the tutor. In the exam, with no AI for anyone, students who had used the open AI did 17% worse than those who never had the tool; the tutor group lost nothing. The authors describe AI being used as a crutch. [6]
With the same tool, you can learn or just deliver
How to read: on the left, students’ grades indexed to 100 for the group without AI. On the right, the range of test scores depending on how programmers used the assistant.
High school maths students
in the exam for those who had open AI
During practice, with the tool
In the exam, no AI for anyone
Calculated index · no AI = 100 · scale 0 to 250
Nearly 1,000 students · one school in Turkey
PNAS · 2025 [6]Programmers facing an unfamiliar library
out of 100 for those who question the tool, against 24–39 for those who delegate
Test score without AI, range by usage pattern
Scale 0 to 100 · no-AI group average as reference
12 people per family of patterns · 3 patterns each
Shen & Tamkin · 2026 [4]Takeaway: the gap is less between “with” and “without” AI than between two ways of using it.
It is also what I experience with this blog, the apps I build and my CRM. I know the need, but I did not have the technical skills to build these tools. With AI, I move forward while gradually understanding the back end, databases and GitHub, and what makes me understand is questioning the solutions I am given rather than accepting them as they are. I have not become a developer, but I think more like a product owner, with one advantage over a beginner: I already know my job, so I know what to question.
The limits of these numbers
These results rest on fragile foundations, and I would rather say so before drawing anything from them. In the Trio experiment, each family of patterns covers twelve people, most of them developers with several years of experience, and understanding is measured an hour after the exercise. The Turkish study covers a single school and mathematics. The customer support study covers one company and one job.
On employment, the US data are descriptive, the French index measures potential exposure, and the Portuguese survey relies on volunteer respondents. My 70 positions describe one portfolio, not a market. Finally, the tools change faster than the studies: METR had to redesign its protocol because some developers now refused to work without AI. [7]
None of this proves that AI reduces junior hiring, or that one way of using it guarantees learning. These are signals pointing in the same direction, nothing more.
My working hypothesis
With those limits in mind, here is where I stand. It is not a conclusion; it is a position I want to test against yours.
First hypothesis: AI does not decide on its own what a beginner learns. It depends on the junior, and above all on how they question the tool. Someone who asks why, checks and challenges an answer learns while producing. Someone who delegates delivers, sometimes faster, but will not be able to take the work over alone. Critical thinking becomes a starting skill, and no longer just a quality expected of seniors.
Second hypothesis: that critical thinking develops better within a framework, and that framework cannot rest on a single player. The tutor in the Turkish study did not hand out answers; it made students search. In a company, that role falls to people and to assignments designed for learning. Consulting firms can run structured junior hiring programmes, schools and companies can offer more internships, and end clients need to commit more, by setting aside part of their assignments for people who are still learning.
On internships, the comparison with France interests me. In the paths I knew there, internships punctuate a degree: students discover a company before finishing their studies, sometimes several times, then return to class with that experience, under a formal agreement between student, school and employer. [8] Portugal also has internships built into degrees, but the most visible public scheme, Estágios +Talento, run by the national employment service IEFP, has a different purpose: it helps already qualified young people into work and explicitly excludes curricular internships. [9] One builds experience during studies, the other supports entry after them.
Several of my contacts in large French banks have told me they prefer the French system on this point; that is their view, which I am reporting. In France I also saw junior hiring campaigns run with engineering schools, where the first role was prepared with the teams. In Portugal, I see academies and small firms take that bet. But across my 70 positions, the path to an assignment with a large client remains rare for someone with zero to two years of experience.
At the start of a change
If a young graduate asked me for advice tomorrow, I would tell them to use every tool available to them as much as they can, but to question hard every solution they are offered. For me, the key is critical thinking, and the ability to ask the right questions.
My generation lived through the arrival of the internet. Looking back, I can see how much it changed the way I work, keep informed and meet people. In its early days, I could never have imagined the place it would take. AI gives me the same feeling of standing at the beginning of a change whose scale is still hard to make out.
We can already see effects on how young people are hired, and the first studies on learning with AI say something fairly simple: the tool amplifies what you bring to it. I wonder what beginners will build with it, which careers they will choose and what frameworks companies will agree to offer them (the short answer is that it depends on the junior, which is precisely why the question interests me).
What do you think?
- If you started your career less than three years ago, has AI mostly helped you understand, or mostly helped you deliver?
- On the company or client side, what stops you today from giving a supervised assignment to someone with zero to two years of experience?
- How do you assess a junior candidate’s critical thinking when AI can answer almost every technical question in an interview?
Sources and method
- Palladino et al. · Exposure of occupations to generative AI and the labour market in FranceLIFT Working Paper 2026-02, October 2026 (in French). Administrative labour-flow data up to March 2026. Permanent-contract (CDI) hires of 21–29 year-olds by occupational exposure: −28.3% and −7.5%. Banque de France summary (in French).
- Landing.Jobs & Damia · Tech Talent Trends Report 2026Summary published on 21 May 2026. Survey of tech professionals in Portugal combined with recruitment interview data. Share of professionals with less than three years of experience: 12%. Self-selected sample. Coverage in Human Resources Portugal (in Portuguese): demand concentrated on senior profiles.
- Brynjolfsson, Chandar & Chen · Canaries in the Coal Mine?Stanford Digital Economy Lab, August 2026 version. ADP payroll data up to June 2026. Employment of 22–25 year-olds in exposed occupations: 19% below the trajectory of less exposed occupations. Decline concentrated where AI automates; employment flat or rising where it complements work. Descriptive indicators, some pre-existing trends, effect attenuated when controlling for education.
- Shen & Tamkin · How AI Impacts Skill FormationRandomised experiment with 52 programmers learning Trio, GPT-4o assistant during the exercise only. Figure 6: scores on the test without AI. Section 6 and figure 11: six usage patterns, scores of 24–39% for those who delegate and 65–86% for those who question, twelve people on each side.
- Brynjolfsson, Li & Raymond · Generative AI at WorkQuarterly Journal of Economics, 2025, 140(2). 5,172 customer support agents. Authors’ full text, section 4.2.3: about 2.5 resolved issues per hour reached in two months with the assistant, in eight to ten months without. Section 5.2: gains maintained during outages, stronger for agents who followed the suggestions.
- Bastani et al. · Generative AI without guardrails can harm learningPNAS, 2025, 122(26). Field experiment with nearly 1,000 high school students in Turkey, GPT-4. Practice: +48% (open AI) and +127% (tutor). Exam without AI: −17% for open AI; the negative effect largely avoided with the tutor.
- METR · We are Changing our Developer Productivity Experiment Design24 February 2026. Selection bias linked to refusals to work without AI, and measurement difficulties with multiple agents.
- Service Public · Student internships in a professional settingFrench framework for internships built into a degree and the internship agreement (in French). Accessed 10 October 2026. The preferences of banking contacts are reported by the author, without a representative survey.
- IEFP · Estágios +Talento regulationsPoints 1.3 and 2: entry into work for qualified young people registered with the IEFP; point 2.2 excludes curricular internships (in Portuguese). Accessed 10 October 2026. University of Lisbon: curricular internships.
How to read these data
Field observation: the author’s count of the 70 positions open in his portfolio in Portugal in October 2026, two of them for zero to two years of experience. The decline in junior requests is a qualitative observation, without a historical series.
France: permanent-contract hires of 21–29 year-olds, private sector, seasonally adjusted and smoothed series, November 2022 to March 2026. Exposure index built from the skills listed in France Travail’s occupational directory, assessed by a language model. Age is not seniority. No causal effect is estimated.
United States: the 19% figure compares the employment trajectory of 22–25 year-olds in the two most exposed quintiles with that of the three least exposed. The ADP sample over-represents large firms and exposed occupations; the gap is less pronounced in national surveys.
Student indices: no-AI group set at 100, based on the published changes (+48%, +127%, −17%). For the tutor in the exam, the gap with the no-AI group is not significant; the displayed value is rounded to 100.
Trio usage patterns: ranges of average scores by pattern, for the 24 AI-assisted participants whose recordings could be analysed. Participants chose their usage freely; the comparison describes an association, not an effect.
Customer support: the indicator measures output with the tool in one company. Lasting learning is inferred from outage periods, with the caveats expressed by the authors.