Optimising the effectiveness of large language models

Thursday 14 May 2026

The contemporary enthusiasm surrounding large language models has produced a curious and often underexplored phenomenon: the widening cultural divide between the people who build these systems and the professional communities expected to use them. The public conversation surrounding artificial intelligence frequently assumes that once the engineering challenge has been solved, practical adoption follows naturally. Yet in reality the greatest obstacles often emerge not from the technology itself but from the human difficulty of cooperation between radically different intellectual cultures.

Software engineers, machine learning specialists, lawyers, doctors, financiers, military planners, journalists and public administrators frequently speak entirely different professional languages. Each group possesses its own assumptions about evidence, reliability, hierarchy, responsibility and risk. Large language models therefore sit at the intersection of technical capability and institutional culture, and it is at this intersection that many projects begin to fail.

The problem is not merely technical illiteracy amongst non-technical professionals. Equally significant is the tendency of technical specialists to misunderstand the practical realities of professional work outside software development. Many information technology experts approach optimisation as an engineering exercise. Yet most professions are not structured like engineering environments. Legal advice, medical diagnosis, diplomatic negotiation, journalism and military command involve ambiguity, incomplete information, emotional judgment and institutional accountability. The assumption that these domains can simply be “automated” often reflects a profound misunderstanding of how professional decision-making actually functions.

A lawyer, for example, rarely produces advice by merely identifying the formally correct legal rule. Legal work frequently involves predicting how judges may react emotionally to arguments, estimating the political atmosphere surrounding litigation, interpreting contradictory evidence, managing client psychology and balancing commercial objectives against legal exposure. An experienced litigator may reach conclusions that are difficult to explain in purely algorithmic terms because the reasoning incorporates years of tacit observation accumulated through practice. When software engineers attempt to model such work purely as a data-processing problem, professional users often become frustrated by the resulting systems.

The reverse misunderstanding is equally common. Many professionals outside information technology possess unrealistic expectations of what large language models can achieve. They frequently regard the models either as miraculous oracles or as fundamentally unreliable parlour tricks, without understanding the probabilistic architecture that governs their behaviour. This creates severe communication difficulties. A senior executive may demand perfect factual accuracy from a system fundamentally designed around statistical language prediction, while simultaneously refusing to allocate the resources necessary for proper human supervision or domain-specific training.

The consequence is mutual distrust. Engineers perceive professional users as irrational, conservative and technologically illiterate. Professionals perceive engineers as naive, arrogant and detached from operational reality. Large language model projects often collapse in this atmosphere of reciprocal incomprehension.

One of the greatest structural difficulties lies in the concept of hallucination. Within the machine learning community, hallucination refers to the generation of plausible but false outputs. Engineers may regard hallucination rates as an optimisation parameter to be gradually reduced through improved architectures, retrieval systems or reinforcement learning. But for many professions, even rare hallucinations are institutionally intolerable.

A physician cannot casually provide incorrect pharmaceutical advice merely because the system is statistically accurate ninety-nine percent of the time. A military officer cannot rely upon a targeting analysis system that occasionally invents battlefield data. A solicitor cannot submit fabricated legal citations to a court. The professional consequences of even isolated errors may be catastrophic. Hence the professional community often evaluates large language models according to standards fundamentally different from those used by software developers.

This creates a profound mismatch in incentives. In the culture of software engineering, iterative improvement and deployment are regarded as normal. Products are released in imperfect form and refined over time. In medicine, aviation, law or military operations, such experimentalism may be unacceptable because institutional trust depends upon reliability rather than innovation velocity.

The difficulty becomes particularly acute when organisations attempt to integrate large language models into existing bureaucracies. Large institutions are frequently built around procedural accountability structures developed over decades or centuries. Every decision-maker possesses formally defined responsibilities. Yet large language models blur responsibility in dangerous ways. If an artificial intelligence system drafts an inaccurate report subsequently approved by a human supervisor, who bears liability? The engineer who designed the model? The executive who authorised deployment? The employee who failed to detect the error? The institution itself?

Such questions are not merely theoretical. They strike at the foundations of professional legitimacy. Many professions derive social authority precisely from the notion that trained individuals exercise accountable human judgement. Large language models threaten to diffuse that accountability into opaque technological systems whose reasoning processes even their creators only partially understand.

Moreover the technical culture surrounding artificial intelligence frequently underestimates the importance of institutional memory and human relationships. Professional environments often operate through informal trust networks developed over many years. A diplomat may know that a particular foreign official responds better to conciliatory language than to direct criticism. A journalist may recognise subtle indicators that a source is unreliable. A military commander may understand the emotional resilience of individual subordinates under stress. Such knowledge is difficult to encode into datasets because it exists partly within human social experience rather than explicit documentation.

Consequently many attempts to optimise professional workflows through large language models encounter resistance not because professionals oppose innovation, but because they recognise dimensions of their work invisible to technical designers.

Another major challenge lies in language itself. Information technology specialists frequently use terminology incomprehensible to external professionals: transformer architectures, tokenisation, embeddings, retrieval augmentation, parameter quantisation, reinforcement learning from human feedback. Meanwhile professional users employ specialised vocabularies equally obscure to engineers. Legal doctrine, financial regulation, military procurement, medical ethics and journalistic standards each contain their own conceptual universes.

When interdisciplinary teams fail to establish a shared vocabulary, cooperation deteriorates rapidly. Meetings become exercises in parallel monologues rather than meaningful collaboration. Engineers present technical possibilities detached from operational realities, while professional users articulate requirements in ways insufficiently precise for implementation. Entire projects can consume millions of dollars while the participants never fully understand one another.

The economic structure of the artificial intelligence industry compounds these problems. Venture capital incentives reward rapid scaling, publicity and disruption. Yet professional integration often requires patience, domain-specific adaptation and slow institutional trust-building. The result is a constant pressure to exaggerate capabilities while minimising discussion of limitations. Professionals subjected to unrealistic marketing claims frequently become cynical after encountering the actual systems.

This problem is especially visible in journalism. Large language models can already produce grammatically coherent articles, summaries and translations. Yet journalism is not merely the production of syntactically correct text. It involves source evaluation, political judgement, moral responsibility and contextual understanding. A journalist covering corruption in wartime Ukraine, for example, may rely upon subtle personal relationships and assessments of danger that no language model can independently replicate. When technology companies imply otherwise, professional journalists often interpret such claims as evidence that engineers fundamentally misunderstand the profession they seek to transform.

The geopolitical dimension further complicates matters. Large language models are increasingly becoming instruments of state power, military capability and economic competition. Governments wish to exploit them for intelligence analysis, propaganda management, cyber operations and industrial productivity. Yet the technical specialists developing these systems may possess very different ethical assumptions from the political authorities deploying them. Conflicts therefore emerge not merely between professions but between entire institutional philosophies.

One sees this tension particularly clearly in military applications. Engineers frequently prioritise optimisation efficiency, data integration and autonomous responsiveness. Military institutions prioritise command hierarchy, predictability and rules of engagement. A system optimised for rapid adaptive decision-making may simultaneously undermine the clarity of human command responsibility upon which military discipline depends.

The educational system has not yet adapted adequately to these interdisciplinary realities. Most software engineers receive little formal training in law, politics, ethics or organisational psychology. Most professional degree programmes similarly provide limited technical education regarding artificial intelligence systems. Consequently both sides approach cooperation with insufficient conceptual tools.

Perhaps the greatest danger is that institutions may respond to these difficulties not through deeper cooperation but through superficial imitation. Organisations increasingly create “AI strategies” largely for reputational reasons, deploying language models in symbolic ways that produce little substantive benefit. Employees may quietly ignore the systems while executives publicly celebrate technological modernisation. Such performative adoption creates an illusion of progress while obscuring the unresolved structural problems beneath.

Nevertheless despite these difficulties, genuine cooperation remains possible. Indeed it may become one of the defining professional skills of the twenty-first century. The most successful large language model deployments are often those in which engineers and domain experts work together continuously rather than sequentially. Instead of engineers building systems in isolation and presenting them to professional users afterwards, collaborative development allows institutional knowledge and technical understanding to evolve together.

This requires humility from both sides. Engineers must recognise that many professions contain forms of tacit knowledge resistant to purely computational modelling. Professional users must similarly accept that artificial intelligence systems can augment human capability even if they cannot fully replicate human judgement. Effective cooperation depends less upon technological perfection than upon realistic mutual understanding.

The future effectiveness of large language models may therefore depend less upon breakthroughs in computational power than upon the social capacity of highly specialised professional cultures to communicate with one another. Artificial intelligence is not merely an engineering project. It is an institutional project, a linguistic project and ultimately a civilisational project. The challenge is not simply teaching machines to imitate human language. It is teaching humans from radically different intellectual traditions to understand one another well enough to use those machines wisely.