Future Directions in Intelligence Enhancement

Single-gene disorders and complex traits are different problems

A pathogenic variant can sometimes be a major cause of a rare disorder. That gives researchers a comparatively clear target: correct, replace, silence or compensate for a defined molecular error. Even then, timing and delivery matter; preventing damage may be easier than reversing a brain that has already developed under the condition.

Intelligence is different. Genome-wide studies indicate that cognitive ability is highly polygenic: many variants contribute, each with a small statistical association, and their effects unfold within environments.² A polygenic score summarizes weighted genetic associations and may estimate relative propensity within a studied population; it is not a molecular blueprint, a diagnosis or a forecast of one individual’s ceiling.

This complexity does not make genes unimportant. It makes simplistic promises irresponsible. Biological inheritance helps shape cognitive differences, while development and sustained learning build the functioning person. Respect for intelligence therefore includes respect for both inherited variation and the years of effort through which knowledge, judgment and expertise are formed.

Communication without repaired neurons

A brain–computer interface can record patterns associated with intended speech or movement, decode them with software and control an external device. This may restore a route for action even when it does not heal the injured pathway. That is not a lesser accomplishment: communication, mobility and self-direction can transform a life.

China’s regulator authorized a minimally invasive extradural BCI in 2026 for a narrow group of adults with stable cervical spinal-cord injury to control an assistive glove.⁵ That authorization is significant because it moves a restorative interface into defined clinical use. It should not be reported as approval of a general brain implant, communication device or intelligence enhancer.

In a 2026 home study, one participant with amyotrophic lateral sclerosis used an implanted speech BCI for 19 months, with average decoding around 56 words per minute and caregiver-supported setup.⁶ It was a remarkable single-participant demonstration, not a consumer product and not evidence that thoughts can be read freely. These systems infer trained intentions from particular signals under particular conditions.

Non-invasive does not mean trivial

Transcranial magnetic stimulation (TMS) and transcranial direct-current stimulation (tDCS) affect the brain without surgery. Some devices are authorized for specific disorders. For example, in late 2025 the FDA approved a prescription tDCS system for a defined subset of adults with moderate-to-severe major depressive disorder.¹⁰ This says nothing by itself about use in healthy people, learning, creativity or general intelligence.

Stimulation studies are sensitive to electrode or coil placement, intensity, timing, individual anatomy, task and blinding. Small changes in experimental design can change the result. Consumer devices therefore deserve the same basic questions as medicines: what exact endpoint, compared with what, in whom, for how long, with what adverse-event monitoring?

Orexin shows why context matters

Narcolepsy type 1 involves loss of orexin-producing neurons. In August 2026, the U.S. FDA approved oveporexton, the first orexin-2 receptor agonist for adults with this condition, following two 12-week randomized trials.¹⁴ That is an important form of biological restoration: supplying a signal a disease has taken away. It is not evidence that a healthy person can safely eliminate sleep or gain supernormal cognition. Insomnia was among the common adverse reactions.

This distinction protects scientific progress from hype. A treatment can be revolutionary for the people who need it without becoming a universal “smart drug.”

Dopamine precision is not yet cognitive precision

Receptor selectivity is attractive because broad dopamine activation can produce unwanted effects. But a more selective mechanism still has to demonstrate meaningful outcomes. Mevidalen, a dopamine D1 positive allosteric modulator, reached phase II testing in dementia with Lewy bodies and did not meet its cognition endpoints; higher doses also raised safety concerns.¹⁵ Negative and neutral trials distinguish plausible mechanisms from treatments that meaningfully help people.

Plasticity is an opportunity, not intelligence in a bottle

So-called psychoplastogens are studied for their ability to promote structural or functional plasticity. Some non-hallucinogenic analogues have produced cellular and animal findings. Tabernanthalog, for example, has preclinical evidence—not established human cognitive benefit.¹⁶

A rise in BDNF, a dendritic-spine change or a “plasticity” label is a mechanism or biomarker. It does not prove better memory, reasoning or wisdom. Plasticity can enable learning and recovery, but experience determines what is learned; change can be adaptive, neutral or harmful.

Similarly, microdosing should not be marketed as settled cognitive enhancement. A preregistered 2026 meta-analysis of 14 studies found no overall cognitive benefit and a small reduction in cognitive control, with substantial limitations in the underlying literature.¹⁷

A revealing classroom result

In a 2025 field experiment involving nearly 1,000 secondary-school mathematics students, unrestricted GPT assistance improved performance during practice. Yet when access was removed, those students scored about 17% below the control group on the unaided examination. A purpose-built tutor that used staged hints and encouraged students to attempt problems avoided that measured learning loss, although it did not produce a significant unaided advantage over the control.²⁰

The lesson is not that all chatbots make people less intelligent. It is narrower and more useful: a tool can improve the work visible on the screen while leaving less knowledge inside the learner. Guardrails that preserve effort matter.

Another randomized study, in a Harvard introductory physics course, found higher immediate post-test performance and shorter median study time with a carefully engineered AI tutor than with an active-learning class.²¹ But the experiment covered two lessons, used immediate tests and did not measure long-term retention. Its tutor contained expert-prepared material, staged questioning and course-specific design. It is evidence that a well-built tutor can help—not that a general chatbot can replace teachers.

Teachers become more important, not less

A teacher does more than deliver information. Teachers select worthwhile goals, recognize misconceptions, judge when struggle is productive, connect knowledge with consequence, cultivate discussion, protect vulnerable learners and assess what a person can do independently. AI can help generate examples, translate material, create practice or prepare a first feedback pass. The educator remains the designer, witness and accountable professional.

UNESCO’s guidance calls for privacy protection, age-appropriate use, ethical validation and a human-centred approach to generative AI in education.²² That is more than a compliance exercise. Students need spaces to think without surveillance, alternatives when a system is inaccessible, and human appeal when automated tools influence opportunity.

At work, the frontier is jagged

Controlled and field studies show genuine productivity gains in some bounded tasks. In an experiment with 453 professionals, generative AI reduced time on short occupational writing tasks by 40% and increased independent quality ratings by 18%.²⁴ In a study of 758 consultants, AI users were faster and performed better on selected tasks inside the model’s capability frontier—but were less likely to reach the correct answer on a task deliberately placed outside it.²⁵

Deployment to 5,172 customer-support agents increased issues resolved per hour by 15% on average, with larger gains among less experienced workers.²⁶ This suggests that systems can diffuse some expert practices. It does not mean every occupation, expert or workflow receives the same benefit.

A meta-analysis of 106 experiments found that human–AI systems generally outperformed humans alone, yet on average performed worse than whichever agent—human or AI—was best by itself.²⁷ “Human plus AI” is therefore a design problem, not an automatic victory. Teams need calibrated trust, clear handoffs and a way to recognize when the machine has crossed from helpful to confidently wrong.

Tasks change before whole occupations disappear

The International Labour Organization’s refined 2025 index estimated that one in four workers worldwide is in an occupation with some exposure to generative AI, while 3.3% of global employment falls in its highest-exposure category. Exposure was greater in high-income economies and, in the highest category, among women. The ILO judged task transformation more likely than complete replacement because most occupations combine automatable and human-dependent work.²⁸

Exposure is not a layoff forecast. It describes technical overlap with tasks, not adoption, economic demand, bargaining power or a fixed timeline. But productivity is not job security either. Organizations and governments must measure who receives the gains, who bears transition costs and whether new roles provide real development rather than permanent oversight of machines.

One danger deserves early attention: if employers automate the junior tasks through which people learn a profession, they may gain short-term efficiency while weakening the supply of future experts. Apprenticeship must be redesigned, not deleted. New entrants need supervised practice, explanation, increasing responsibility and regular opportunities to perform core work without AI.

What a higher score can honestly mean

A higher score on a valid cognitive assessment is evidence of stronger performance in the abilities that assessment actually measures. Those abilities are real and consequential. In the right context, they can support faster learning, more complex reasoning and the capacity to solve problems that require unusually deep abstraction.

That statement should not be weakened into “scores mean nothing.” Nor should it be stretched into “one score measures every valuable property of a person.” No IQ test fully captures acquired knowledge, intellectual courage, creativity, moral judgment, practical wisdom, motivation, consciousness or the future contribution someone will make. Measurement earns respect by remaining precise.

Intelligence is not a toy purchased from a shelf. Its development reflects inherited differences, physical brain development, health, education, practice, opportunity and a lifetime of accumulated learning. Expertise is especially costly: it can embody decades of attention, correction, memory, sacrifice and experience. When a knowledgeable person is lost, neglected or prevented from contributing, tools and databases do not automatically recreate the organized understanding that existed in that mind.

Why stepping away can help

Groups can focus attention, but they can also anchor it: the first confident suggestion can change what everyone else considers, conversation can interrupt unfinished associations, and social approval can reward familiar directions. The same effect can occur when a person asks AI for ideas before listening to their own mind.

Experimental work on creative incubation found that an undemanding task associated with more mind wandering improved later performance on previously encountered creative problems compared with demanding activity, rest or no break.³⁰ This does not mean all mind wandering is productive or that loneliness improves creativity. It supports a modest principle: mental space and temporary release from focused control can sometimes help an unfinished problem reorganize.

Research on social influence and creativity likewise shows no simple “groups are good” or “groups are bad” verdict. Evaluation pressure, conformity, production blocking, diversity, dissent and the sequence of individual and collective work all matter.³¹ A strong design often lets people generate independently first, then compare, combine and challenge ideas together.

Global rules are beginning to catch up

In November 2025, UNESCO’s Member States adopted the first global normative framework on the ethics of neurotechnology. It emphasizes dignity, mental privacy, affordability, explicit consent and transparency; it warns against non-therapeutic use in children and against workplace monitoring that profiles employees.³² The recommendation is not a self-enforcing global law, but it gives governments and institutions a common direction.

The Council of Europe has separately examined neural data through the lens of Convention 108+, emphasizing protections for mental privacy while permitting responsible scientific progress.³³ The OECD’s recommendation on responsible innovation in neurotechnology likewise calls for safety, inclusivity, deliberation, stewardship of personal brain data, oversight and anticipation of misuse.³⁴

AI governance increasingly reaches education and employment too. Since 2 February 2025, the European Union’s AI Act has prohibited emotion inference in workplaces and educational institutions except for uses intended for medical or safety reasons. It also classifies certain systems that influence educational or employment access as high-risk; the principal obligations for these Annex III systems apply from 2 December 2027.³⁵ These categories recognize an essential fact: a model that grades, hires or monitors can shape a life even if it never touches the brain.

Progress has a physical footprint

AI is not weightless: it runs through physical data centers that require electricity and infrastructure. The International Energy Agency’s 2025 report models rapidly rising data-center electricity demand while also examining how AI may improve energy systems.³⁶ Biomedical technologies also carry laboratory, manufacturing, maintenance, supply-chain and disposal footprints. Responsible enhancement counts the full ledger: clinical benefit, labor, environmental burden and who lives near the infrastructure.

A durable practice for individuals