BY Bella Delmonte ’27
Staff Writer
Sentient robots once seemed like a figment of 1980s sci-fi; however, new hearing and speech abilities for ChatGPT announced by OpenAI could potentially introduce this concept to the real world.
McKinsey & Company has described 2023 as the “breakout year” for generative artificial intelligence. Generative AI is a branch of AI that learns from and imitates vast amounts of data to create content based on inputs or prompts, per the Harvard University website. Generative capabilities include text, images, music, videos, code and more, the same website described.
According to their website, ChatGPT is a generative AI program created by the Microsoft-backed startup OpenAI. According to OpenAI, the program’s dialogue format allows it to answer initial and follow-up questions, admit its own mistakes, challenge incorrect assumptions and reject inappropriate requests.
To some extent, generative AI may now even mimic human conversation, as OpenAI announced in late September that ChatGPT will soon include hearing and speech capabilities. In the same announcement, they said they will soon begin to release the vision features of GPT-4, allowing the program to perceive and analyze visual inputs.These capabilities could ultimately lend themselves to assistive technology for the visually-impaired.
Since the original ChatGPT’s release last November, it has raised some eyebrows in schools. The New York Times reported that some administrators and educators have amended honor codes and teaching practices in a fight against the machines in the battle for honesty, integrity and quality in education.
Though not an overwhelming majority, a notable portion of students may be turning to these programs for help. A survey by BestColleges shows that out of 1,000 college students surveyed this March, roughly 22% admit to using or having used generative AI for their schoolwork.
According to The New York Times, ChatGPT is a common suspect for students turning to AI assistance for schoolwork. Reuters reported that its popularity surged to 100 million users in the first two months of its release, and its writing capabilities — though suggested by OpenAI to be mediocre based on its Advanced Placement English scores — are a convenient escape for students unable to complete their work or who just can’t be bothered.
While the educational perspective of the issue is a troubled one, the financial prospects for developers of generative AI are looking bright, Reuters reported. According to Reuters, OpenAI expects to reach over $1 billion in revenue in 2023, currently generating more than $80 million each month, compared to $28 million last year.
The Big Technology newsletter suggests that Generative AI has also transformed the job market by increasing demand for engineers who can create, maintain, manage and fine-tune generative AI systems. According to LinkedIn, the demand for AI engineers in 2023 has grown at least twofold, and based on data from Indeed, OpenAI is averse to offering salaries ranging from $95,000 to $578,000.
The sense of importance among builders of AI comes from more than the high demand for their expertise and generous salaries. Thomas Krendl Gilbert, a machine ethicist, asserts in an Algorithmic Bridge article that these engineers think “what they’re doing is ‘magical.’” In the article, he implies that the manufacturers of these systems believe they are bringing humankind closer to the kind of superintelligence once limited to the domain of science fiction.
This “magic” is also underpinned by AI’s potential for good. For example, the Danish startup Be My Eyes is collaborating with OpenAI, using GPT-4 to transform visual accessibility, the OpenAI website detailed. As explained by OpenAI, BME creates technology for those who are blind or visually impaired by connecting them with volunteers who help with hundreds of daily life tasks.
With the new visual input capability of GPT-4, BME began developing an AI feature within the Be My Eyes app that can generate the same level of context and knowledge as a human volunteer, according to OpenAI. BME CEO Michael Buckley has stated, “In the short time we’ve had access, we have seen unparalleled performance to any image-to-text object recognition tool out there.”
Despite the innovative prospects of AI, Gilbert describes it in another Algorithmic Bride article as “mystic exploration without understanding.” The development of many AI systems is dubious as it is achieved mainly through trial and error, and developers do not focus on scientific processes or theory so much as quickly achieving new goals.
Consumers may want to employ some healthy skepticism for these programs, as TechCrunch suggested in an article covering the lingering flaws of GPT-4V, the visual component of GPT-4.
Consider the details of a system card released last week by OpenAI covering GPT-4V, currently in research preview. The program, as described by OpenAI, intends to enable users to instruct GPT-4 to analyze image inputs provided by the user. However, the company admits to some faulty aspects of the feature.
Though it can occasionally correctly identify toxic foods such as poisonous mushrooms from images, it has misidentified substances like fentanyl, carfentanil and cocaine from images of their chemical structures, OpenAI explained. Their system card also reports GPT-4V’s facial recognition systems having disparate performance based on race and the program being prone to stereotyping members of different races.
According to The New York Times, while machines learn to execute tasks through data analysis, they sometimes develop the types of harmful and unexpected behavior patterns explained in the system card. Through human feedback, however, OpenAI says they are attempting to reduce the flaws of GPT-4 and 4V. The New York Times suggests that human cognitive reasoning skills serve as an important check to reduce the misinformation, bias and other toxic information produced by generative AI.
However, this feedback may have damaged other aspects of the program, as a recent study from researchers at Stanford University and the University of California, Berkeley, shows that the accuracy of ChatGPT in mathematics, coding and reasoning skills has diminished in the past several months, potentially due to this human feedback. Finding an equilibrium between programs being “too human” and “not human enough” may require a different approach.
Though engineers and consumers alike are finding ways to incorporate generative AI into their lives, the future role of AI remains foggy in the opinion of some scientists, such as Gilbert, who suggests AI is going nowhere fast and “can’t really bring any valuable insights.”
However, according to Cambrian AI Analyst Alberto Romero, AI could be only a piece of the puzzle. Romero points out that, like alchemy giving way to chemistry, AI might be the antecedent of a new science yet to be unveiled.