The Ghost in the Inbox: How to Tell Whether You Are Talking to a Human or a Machine

By Matthew Parish, Associate Editor

Thursday 18 June 2026

The rise of large language models has quietly transformed electronic communications. A few years ago, when one received an email, a text message or an instant message, there was a strong presumption that another human being had written it. That assumption is no longer safe. Today a growing proportion of written communications are generated, partially generated or heavily assisted by artificial intelligence systems. Some messages are drafted by humans and polished by machines. Others are written entirely by software. Increasingly, systems such as Claw and similar computer-use agents can not only generate text but also interact with software, browse websites, read emails, complete forms and carry out extended chains of actions with little or no human supervision.

This creates a profound epistemological problem. When reading a message on a screen, how can one know whether there is a person at the other end or merely a statistical language model producing plausible sentences?

The answer is becoming increasingly difficult.

The End of the Turing Test

For decades artificial intelligence researchers referred to the “Turing Test”, proposed by the British mathematician Alan Turing in 1950. The idea was simple. If a machine could converse with a human without revealing its non-human nature, then the machine could be said to exhibit a form of intelligence.

For most of the history of computing, this was largely theoretical. Computer programs produced awkward responses, misunderstood context and generated obviously mechanical language. Identifying a machine was relatively straightforward.

Large language models have changed this entirely.

Modern systems can write persuasive emails, conduct lengthy conversations, imitate professional styles, generate humour, express apparent empathy and maintain contextual coherence across thousands of words. They have become sufficiently sophisticated that many people interact with them daily without necessarily recognising their involvement.

The problem is compounded by agentic systems such as Claw. Rather than merely generating text when prompted, such systems can execute tasks autonomously. They can search databases, analyse documents, compose correspondence and respond to incoming messages. A human recipient may therefore be engaged in what appears to be an ordinary conversation while the other side is, in reality, a software system conducting an extended workflow.

The Strange Perfection of Machine Writing

One of the first clues lies in linguistic perfection.

Humans are imperfect communicators. They make typographical errors. They omit words. They occasionally contradict themselves. They become distracted, tired or emotional.

Large language models often display a suspicious degree of consistency.

A message that is flawlessly structured, grammatically impeccable and endlessly polite may not have been written by a machine, but the possibility should be considered.

This becomes especially noticeable over long exchanges.

Human correspondents tend to vary their style depending upon mood and circumstance. One day they may write formally; the next day they may be brief and hurried. Their vocabulary fluctuates. Their attention wanders.

Artificial systems often maintain a remarkable stylistic consistency. Every message may have the same cadence, the same sentence structure and the same balanced tone.

The result can feel oddly artificial even when the language itself appears natural.

Excessive Helpfulness

Another characteristic of large language models is their tendency towards excessive helpfulness.

Humans frequently answer only the question that was asked. Machines often answer the question plus three related questions, provide a summary, offer suggestions and conclude with additional recommendations.

The response may seem almost unnaturally eager.

For example, a human asked for a train timetable might simply provide the departure time. A language model may provide the departure time, alternative routes, historical context, travel advice and a discussion of weather conditions at the destination.

This tendency arises because language models are trained to maximise perceived usefulness.

Ironically, this often makes them easier to identify.

The Absence of Genuine Memory

One of the most revealing tests involves personal continuity.

Humans possess experiences, memories and emotional histories. Their communications are shaped by events that have happened to them.

Language models do not possess memories in the human sense. They may maintain contextual information during a conversation, but they lack lived experience.

Questions involving shared experiences can therefore become revealing.

A human colleague might remember an awkward meeting six months earlier, an inside joke or a private conversation. A machine will either lack such knowledge entirely or produce generic responses.

As agentic systems become integrated with databases, calendars and archives, this distinction may become less obvious. Nevertheless there remains a fundamental difference between recalling a stored record and remembering a lived experience.

Speed as a Clue

Response times can sometimes provide hints.

Humans require time to think.

Complex questions may take minutes or hours to answer.

Machines can generate extensive responses almost instantaneously.

A highly detailed reply arriving seconds after a complicated query may suggest automation.

However this indicator is becoming increasingly unreliable. Humans can use artificial intelligence to draft responses rapidly, while automated systems can be configured to delay their replies in order to appear more natural.

What was once a useful signal is gradually disappearing.

Emotional Flatness

Human beings possess emotional irregularities.

They become annoyed. They misunderstand intentions. They react unexpectedly.

Large language models generally exhibit a more uniform emotional profile.

Even when programmed to simulate emotion, the simulation often lacks the unpredictability of genuine human feeling.

Many users report a distinctive sensation when conversing with advanced models. The language appears emotionally appropriate, yet something feels absent. The conversation lacks the subtle irrationalities that characterise human interaction.

A machine may express sympathy perfectly. A human may express sympathy awkwardly but sincerely.

Paradoxically, the awkward response is often the more convincing evidence of humanity.

The Knowledge Boundary Test

Another useful technique involves probing the limits of knowledge.

Humans generally know some things very well and other things poorly.

Language models tend to possess broad but uneven knowledge.

When asked questions close to a person’s area of expertise, a human often provides detailed, experience-based answers. When questioned outside that area, they admit ignorance.

Machines frequently display a different pattern. They may provide confident answers across a vast range of subjects, occasionally introducing subtle inaccuracies while maintaining an appearance of authority.

This phenomenon is particularly evident in professional communications. A lawyer, physician or engineer will often reveal the practical constraints of their profession. A language model may produce theoretically plausible explanations while lacking the tacit knowledge acquired through years of practice.

Large language models also speak multiple languages. Changing the language of communication may result in a fluent response when communicating with a model, without hesitation. Humans do not speak limitless languages fluently.

Claw and the Age of Autonomous Correspondence

Systems such as Claw introduce an additional complication.

Traditional chatbots waited for instructions. Agentic systems can initiate and sustain interactions independently.

Imagine a software agent monitoring an inbox. It reads incoming emails, classifies them, retrieves information from databases, drafts responses and sends replies. A human supervisor may review only a fraction of the correspondence.

In such circumstances the distinction between human communication and machine communication begins to blur.

The sender’s organisation may still be responsible for the content. A human may still stand behind the process. Yet the actual words received by the recipient may never have been read by a person before transmission.

This is likely to become increasingly common in customer service, sales, legal administration, journalism, government and countless other fields.

The question is no longer whether one is talking to a machine.

The question is how much of the conversation is being conducted by a machine.

The Future of Digital Identity

Historically, communication implied presence.

A letter indicated that another person had sat down and written it. A telephone call implied direct engagement. An email generally meant that someone had typed the words appearing on the screen.

That assumption is rapidly eroding.

Over the coming decade, many electronic communications will be generated by hybrid systems combining human oversight with machine execution. Some messages will be written entirely by humans. Others entirely by machines. Most will occupy an ambiguous middle ground.

As a result, discerning whether one is communicating with a bot will become increasingly difficult because the distinction itself is becoming less meaningful.

A more important question may emerge.

Instead of asking, “Am I talking to a machine?”, we may begin asking, “Who is accountable for this message?”

That question strikes at the heart of the matter. Whether words are composed by a human, a large language model or an autonomous agent such as Claw, responsibility must ultimately rest somewhere. The future challenge is not merely recognising artificial intelligence. It is preserving trust, accountability and authenticity in a world where the apparent author of a message may no longer be its true originator.

The age of machine-generated correspondence has arrived. The ghost in the inbox is no longer a hypothetical possibility. Increasingly, it is a daily reality. The difficulty is that the ghost now writes remarkably good English – and any other language you might care to test it on.