Yesterday, C3.ai stock came in above expectations, and Chief Executive Thomas M. Siebel noted, in no uncertain terms, “Our federal revenue grew by more than 100% for the year. The interest we are seeing in our generative AI applications is staggering."
The number of AI companies has increased by over 600% over the last 10 years, and the UK AI market is worth more than £16.8 billion, according to the US International Trade Administration, and is expected to grow to £801.6 billion by 2035. A new Microsoft survey indicates that a stunning 75% of knowledge workers are reportedly using AI for their work, apparently saving over 30 minutes per day.
- 78% of AI users are using their own tools to work.
- Purportedly, 90% say AI makes their work more enjoyable
- 79% of leaders agree AI adoption is critical to remain competitive
- 60% say their company lacks a vision and plan to implement it.
This paints a picture that individuals are quick to take up tools to enhance their workflow but their workplaces aren’t keeping up. AI writing is coming on leaps and bounds, but there are still many tells that a neural network composed the text, not you or I.
Native speakers of English are sometimes offended by its tendencies to use fancy and verbose terminology, but this partly comes down to the trendiness of phrases in culture. For reference, data analysed by Joe Veix for The Outline, indicates that a meme lasts an average of about 4.017 months. A study by Pagel et al. has shown that dominant words rise to the top depending on the frequency of use and push out other terms, and given that memes have a half-life of a few months, the fact that AI is notorious for the use of ‘power’ verbs like fostering, unlocking, cultivating, and harnessing seem dead giveaways.
These linguistic slips-of-the-mask come largely from AI’s inherent weaknesses at this moment in time. For example, the language can often appear choppy, or out-of-context and disjointed. The AI tools predict the best next word, and generally make intelligent choices.
However what if the next most common word or sentence is sometimes out of context or disjointed. Moreover, generative AI companies have used writing widely available on the internet, especially content from online newspapers, published research papers, and government documents. This content is not always ontologically or grammatically accurate, and mistakes are multiplied in the AI workings.
To prevent this problem, these AI tools seem to have algorithms that put in extra transition words between sentences to make the sentences sound more connected. In essence, ChatGPT creates text using prediction, then relies on cohesive devices to link the output to give the impression of coherence. Writers generally group cohesive devices into these six categories:
- Organisational Surface Signals: first(ly), second(ly), third(ly), next, finally…
- Conjunctions: therefore, consequently, as a result, on the other hand, in contrast...
- Summative Transitions: in summary, in conclusion, overall, in other words, to sum up….
- Additive Transitions: also, furthermore, moreover, in addition....
- Prepositions: with, at, by, to, in, for, from...
- Pronouns: he, she, we, they, such...
Reliance on these cohesive devices shrouds the truth, that AI is associating clauses through frequent use, and the connective serves to jam them together. This is why some AI messages will appear choppy, rough, or wholly misguided.
In this article we’ll go through some of the other top AI language red flags to avoid, so you can be aware of the pitfalls of AI and master it.
AI writing has a predilection for certain sentence structures, In the world of…, In the realm of…, You’ll often find a preference for gerunds, like “Mastering the AI language tools…”, or “Upgrading your B2B sales….” or “Unlocking your triggered contacts”.
AI-written text can also follow an unrelenting stream of metaphor and simile, or a symbolism which runs throughout the entire text, dominated by the semantic field of nautical travel, whereby a business is transformed into a vessel, your enterprise casts off from the shore, you sail the seas of sales, and so forth.
LLMs at medium temperature (“a measure of how often the model outputs a less likely token. ) confidently use favourite words like "unleash" and "delve". This links to a larger trend in which words like “tapestry” are used to describe any number of connections, “realm” is used in place of space or sphere, and the view or aspect or anything is described as the “landscape”. May as well be trained on odes and sonnets, elevating the most banal of subject matters to sheer poetry.
AI is infamous for its repetitive use of grammatical structures. My pet peeve begins, “Not only does… but also….” or “it’s not just {this}; it’s {that}”. Another example is the colon-in-the-header - titles like **Navigating the Digital Ocean: The Voyage of B2B SaaS**, whereby the asterisks indicate the heading type in markdown formatting. The colon header is correct, but can’t we have a little more variety? Ironically, the influence of AI can be noted through its sheer determination to use the discourse markers, which are the very thing that identifies itself as overly formal.
Many writers intentionally break grammatical rules for emphasis. Varying sentence length and structure breathes life into text - when read, the variety is interesting, and wills the reader on. AI texts often have no contractions - we’ll’ve shortened parts but AI will not. When the text is too uniform in structure and length, this is a dead giveaway that AI composed the text.
Knowledge of the real world plays a crucial role in both human intelligence and the development of artificial intelligence. Demonstrating intelligent behavior in AI agents heavily relies on their knowledge base. An agent can only accurately respond to input if it possesses some prior knowledge or experience related to that input.
In the scenario in which an LLM is trained, the text and the structures embedded in it represent the data; the topics being written about are not necessarily important to these LLMs. For these researchers, AI will be counted as intelligent if it spit out the right sentence at the right time, if it could manipulate “a representational schema of symbol manipulation.” As Sutskever recently told Johnson in a very telling interview before leaving OpenAI, “The underlying idea of GPT-3 is a way of linking an intuitive notion of understanding to something that can be measured and understood mechanistically, and that is the task of predicting the next word in text.” But does this intuitive notion of understanding entail true consciousness?
In Joan Didion’s words, “All I know about grammar is its infinite power. To shift the structure of a sentence alters the meaning of that sentence as definitely and inflexibly as the position of a camera alters the meaning of the object photographed.” As the current AI models are unable to change the syntax and grammatical orientation of a sentence, it’s unable to wield the infinite power of the couched meaning of words, which are accessed in connection (c.f. De Saussure’s symbiotics).
As Pennebaker notes, measuring language in social psychology is to consider two broad categories: content words and function words. Given AI’s contextual reliance, the system must discern patterns at multiple levels in existing texts, seeing how individual words function and how they tie into the larger passage. In this sense, the word plays a different role in different senses which is only adjacent to the dictionary definition!
Either way, it’s important to be aware of AI language’s nuances and idiosyncrasies. AI is being trained to adopt natural word-choices, but with unideal training data and a programming reliant upon cohesive devices, AI’s ability to write in natural language and be undetectable by humans may be years down the line, or there may be a quantum leap tomorrow!
Today, textual analysis is able to tell us more about ourselves than we know! Research indicates that our self-assessment as Pennebaker wrote back in 2019, textual analysis has been “enabled by advances in computerized text analysis and access to massive archives of
digitized text, we are finding patterns that no one has seen before.” This is increasingly the case - as we learn about AI’s language choice, AI may well teach us about our language choice too!
Comentarios