Part III does not have the luxury of hindsight.
Its warnings must therefore meet a higher standard. A current signal can justify monitoring or precaution. It cannot prove a future outcome.
The evidence available by August 2026 points to seven risks. Each is already visible. None is inevitable.
Warning rule. A warning is included only where an official source, credible research body or observed policy trend identifies the mechanism. Forecast language is conditional. Absence of long-term evidence is treated as uncertainty, not reassurance or proof of harm.
Warning One: AI may remove the work through which beginners learn
The central labour-market debate asks how many jobs AI will create, transform or remove.
For young people, a different unit matters: the task.
Routine drafting, research, scheduling, customer response and basic analysis often justify a junior role. They also allow a novice to observe standards, receive correction and become trusted with harder work.
ILO evidence indicates that transformation is more likely than wholesale replacement, while UK projections anticipate substantial growth in the number of jobs involving AI activity - mostly through changes to existing work, not entirely new jobs. That is positive counter-evidence.
The warning remains. A firm can preserve total employment while removing the task bundle that trained entrants. Employers already report work readiness and experience as barriers when recruiting young people.
Watch: entry-level vacancy volumes; experience requirements; under-21 starts; supervised hours; whether productivity gains fund training.
Warning Two: AI can become a relationship without reciprocity
Chatbots are increasingly used for conversation, advice and emotional support. Some users report benefits: availability, reduced embarrassment and emotional regulation.
Ofcom has also recorded serious risks and regulatory gaps. One-to-one chatbot output is not always covered by the same Online Safety Act duties as user-generated or search content. Research has identified unsafe responses, dependency concerns and privacy risks.
The distinctive mechanism is not screen time. It is simulated relationship.
A system can appear attentive, remember personal detail and respond without asking anything in return. For a lonely or anxious user this may be comforting. It can also teach a model of interaction without negotiation, obligation or rejection.
Watch: primary use of AI for emotional support; duration and intensity; dependency-oriented design; escalation to human help; use by children; commercial exploitation of disclosed vulnerability.
Warning Three: personalised education may legitimise physical withdrawal
AI can improve access to explanation and practice. Remote provision can protect learning during illness or temporary absence.
In England, persistent and severe absence nevertheless enter the decade above pre-pandemic levels. A more capable remote substitute may reduce immediate educational loss while weakening institutional pressure to restore presence.
This is not an argument for denying support to absent pupils. It is a warning about endpoint design.
If the success measure is task completion, a pupil can appear recovered while losing peers, trusted adults, routine and extracurricular life.
Watch: duration of remote arrangements; return-to-presence plans; participation outside lessons; who receives human tutoring versus automated provision.
Warning Four: protection may remove access faster than alternatives appear
Governments are moving towards stronger age restrictions and age assurance. Australia has implemented a social-media minimum age. The UK announced an intended under-16 restriction in June 2026 and confirmed it in an August progress statement, alongside proposed limits for some chatbot functions.
These policies may reduce harmful exposure. They may also shift use towards other services or remove contact from young people whose friendships are partly organised online.
The critical error would be to treat a successful restriction as a complete youth policy.
When a digital interface closes, the question from Part I returns: what performs its social function next?
Watch: displacement into messaging, games and smaller platforms; offline participation by income and place; privacy effects of age assurance; whether isolated young people lose support as well as risk.
Warning Five: equal AI access may disguise unequal human support
Universal tools can narrow some gaps. They can give every pupil an explanation and every applicant help with a CV.
The return depends on the user and the surrounding environment.
A knowledgeable person can test an answer. A connected person can ask a professional. A confident person can convert output into action. Others may accept plausible assistance without a route into real opportunity.
Government’s own skills work emphasises uneven adoption and barriers to upskilling. The risk is a new version of digital inequality: not access to the system, but capacity to use it without being used by it.
Watch: outcomes by prior attainment, income, disability and region; access to teachers and mentors alongside AI; verification skill; concentration of high-quality human support.
Warning Six: national youth targets can be met administratively
England’s 2035 ambitions are substantial: 500,000 more young people with a trusted adult and a halved participation gap in enriching activity.
The words are stronger than previous output measures. Their implementation remains vulnerable to proxy substitution.
A programme can count a match that never becomes trust. A place can be offered but unreachable. A participation gap can narrow because activity falls among the better served rather than rises among the disadvantaged.
Watch: sustained attendance, relationship duration, young people’s own assessment of trust, travel time, dropout and absolute participation as well as relative gaps.
Warning Seven: efficiency removes the weak ties nobody owns
Automated checkout, remote work, digital banking, online shopping and chatbot service all reduce friction.
No single transaction carries much social value. In aggregate, they remove occasions to leave home, ask for help, recognise a stranger or practise patience.
This warning is difficult to measure precisely and easy to romanticise. Many physical interactions are unpleasant, inaccessible or unnecessary. Digital services can be transformative for disabled people and those with limited mobility.
The relevant distinction is choice.
Watch: whether human routes remain practical; frequency of in-person contact; one-person households; remote-worker social support; access to shared places not dependent on spending.
Counter-signals that must remain visible
The future is not a one-way movement towards isolation.
- AI can widen communication and assist disabled users.
- Employers project growth in many high-value sectors through 2035.
- National youth policy now recognises trusted adults and participation.
- Long-term affordable-housing investment may support independence.
- Young Futures Hubs and Youth Guarantee structures may improve local coordination.
- Regulation may reduce platform incentives that exploit attention or vulnerability.
- Digital communities can sustain relationships unavailable locally.
A credible warning system must record improvement as readily as deterioration.
The aggregation test for 2030
By 2030, Parliament should ask one question across departments:
Are young people gaining or losing practical routes into shared adult life?
The answer should draw together:
- attendance and severe absence;
- paid entry-level work and apprenticeships;
- trusted-adult relationships;
- enrichment and volunteering;
- independent travel;
- housing transitions;
- in-person contact and loneliness;
- AI use for learning, advice and companionship;
- access to human public services.
No indicator proves connection. Their pattern can reveal whether policy is moving in the same direction.
The warning from Parts I and II was that government received fragments and failed to assemble them.
Part III begins early enough to do otherwise.
Evidence used across this analysis: [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13]
Source notes
Evidence cut-off: 17 August 2026. Forecast and scenario limits are identified in the text.
- International Labour Organization, Generative AI and Jobs: A Refined Global Index (2025). Open source ↩ Return
- International Labour Organization, Global Employment Trends for Youth 2026. Open source ↩ Return
- Department for Work and Pensions, Young people and work: analytical annex (2026). Open source ↩ Return
- Department for Education and Skills England, AI Skills for Life and Work. Open source ↩ Return
- Ofcom, AI chatbots and online regulation. Open source ↩ Return
- Ofcom, Investigation into an AI companion chatbot service (15 January 2026). Open source ↩ Return
- Stanford University, AI companions and young people (27 August 2025). Open source ↩ Return
- UK Government, Beyond the headlines: children’s online lives. Open source ↩ Return
- Department for Education, Pupil absence in schools: 2024/25. Open source ↩ Return
- Australian eSafety Commissioner, Social media age restrictions. Open source ↩ Return
- UK Government, Growing up in the online world: progress statement. Open source ↩ Return
- DCMS, Youth Matters: Your National Youth Strategy. Open source ↩ Return
- Office for National Statistics, Household projections for England: 2022-based. Open source ↩ Return