The buzz surrounding ChatGPT’s public debut last year was extraordinary, even by the lofty standards of technological innovations. OpenAI’s natural-language processing system has proven itself capable of crafting recipes, generating computer code, and even parodying literary styles with impressive ease. Its latest version has taken things a step further, offering the ability to describe photographs, a feat that has led many to tout it as a revolutionary breakthrough, likened to the invention of the printing press. Yet, it hasn’t taken long for significant flaws to surface. ChatGPT occasionally "hallucinates", fabricating non-factual information that it confidently presents as truth, and when challenged, it often doubles down on these inaccuracies. Additionally, it struggles with basic logical reasoning.
In essence, ChatGPT is not a general artificial intelligence- a sentient, self-thinking machine- but rather a large language model. This designation means it excels at predicting word sequences based on extensive training on vast amounts of text. The specifics of its training sources remain undisclosed by its developer, OpenAI, but the model’s prowess lies in its ability to identify and replicate patterns within that data.
Amidst the surrounding excitement, it’s easy to overlook what is, in fact, a minor miracle: ChatGPT has solved a problem that for years seemed unattainable to engineers- generating human-like language. Unlike earlier iterations of similar systems, ChatGPT can produce coherent text over multiple paragraphs without descending into gibberish. This accomplishment is even more remarkable when you consider the broader scope: ChatGPT can generate not only highly realistic English text but also instantly produce coherent responses in over 50 languages- a precise count that even the model itself cannot confidently provide. When asked (in Spanish) how many languages it can handle, ChatGPT vaguely responds with "more than 50", clarifying that its proficiency in any given language depends on the amount of training data available. Then, when prompted in Portuguese without warning, it delivers a brief bibliography of your columnist, most of which is accurate, though it does incorrectly state the subject and university of study. The language, however, is flawless.
Portuguese, as one of the world’s major languages, is a relatively easy test case. To push the boundaries further, your columnist engaged ChatGPT in Danish, spoken by only about 5.5 million people. Given that much of Danish online content is in English, the training data for Danish must be significantly sparser than for English, Spanish, or Portuguese. The results were factually incorrect, yet expressed in nearly perfect Danish, with only a minor gender-agreement error detected across all tested languages.
ChatGPT, in fact, underestimates its own capabilities. When asked, it lists 51 languages it ca work in, including Esperanto, Kannada, and Zulu. It modestly refrains from claiming to "speak" these languages, preferring to say it "generates text" in them. This is an understatement. When addressed in Catalan- a language not included on the list- it cheerfully responds in Catalan, "Yes, I do speak Catalan- how can I assist you?" Follow-up questions in Catalan pose no challenge, even when asked if it translates answers from another language into Catalan. ChatGPT denies this, asserting, "I don’t translate from any other language; I search my database for the best words and phrases to answer your questions".
Whether this is entirely accurate remains uncertain. ChatGPT has a tendency not only to fabricate information but also to give misleading responses about the conversation it is currently having. It lacks a true "memory" and instead reprocesses the last few thousand words of a conversation as a new prompt. If a conversation has been in English for a while, it might "forget" that a question was asked in Danish earlier and incorrectly claim that it was asked in English. This reveals ChatGPT to be unreliable not only in its knowledge of the word but also in its understanding of itself.
However, these shortcomings should not overshadow the achievement of creating a model capable of convincingly mimicking so many languages, even those with limited training data. For years, speakers of smaller languages have expressed concern about being overlooked by advancing language technologies. Their worries were well-founded, stemming from the lack of commercial incentives for companies to develop products in languages like Icelandic or Maltese, and the relative scarcity of data available to train such systems.
Remarkably, the developers of ChatGPT appear to have surmounted these obstacles. Although it is too soon to predict the full impact of this technology, this accomplishment alone offers a reason for optimism. As machine-learning techniques continue to evolve, they may require fewer resources- in terms of both programming effort and data- than traditionally thought necessary to ensure that smaller languages are not left behind in the digital age.