Looking at the impact the application of artificial intelligence has on almost every facet of our lives and remaining indifferent is a tall order. Its rapid development carries the weight of a double-edged sword, promising transformative changes that will benefit everyone, after a time of uncertainty.
But rather than speculate on the future, let’s ground the discussion around AI in the present reality. What exactly can it do today? More importantly, what shortcomings does it still need to overcome? A balanced take free from unrealistic promises and fearmongering is called for, and this article hopes to provide it.
What AI Currently Does Well
The umbrella term artificial intelligence encompasses complementary yet distinct technologies like natural language processing, computer vision, and agentic AI. Combined and trained well, these technologies can speed up workflows or help advance our understanding in unforeseen ways.
Among these, agentic AI use cases are particularly compelling, as they demonstrate how autonomous AI agents can make real-time decisions, interact with humans, and augment workflows without constant supervision. We’re currently reaping the most benefits from these core capabilities.
Natural language processing
AI is capable of understanding different forms of language, be that human expression through spoken language or machine code. It can interpret, summarize, and generate text. The applications vary, from content creation at scale to the rise of “vibe coding” as a means of assisting both programmers and laypeople in realizing their projects.
NLP enables conversational interaction that helps people learn. This goes beyond retrieving and presenting information from a database; a large language model can simplify or reframe information so it becomes easier to understand. LLMs can also pick up on emotional cues in writing to an extent, allowing them to perform sentiment analysis applicable to marketing, customer service, R&D, and more.
Pattern recognition
A standard method of teaching an AI is to provide it with data and assign value to it. For example, medical experts have had great success in feeding AI images of malignant tumors and other anomalies. Once it’s been exposed to enough verified data, the AI can extrapolate to new images and identify requested patterns with uncanny precision.
However, structured learning isn’t the only approach. Provided with a dataset and given no limitations, an AI may uncover patterns humans wouldn’t even consider looking for, let alone spot. Banks and credit card companies use this to spot subtle and emerging types of financial fraud. Meanwhile, businesses leverage predictive analytics to account for trends and seasonal fluctuations.
Workflow augmentation
Artificial intelligence exceeds the limitations of older trigger-based automation systems. Specifically, AI agents aren’t bound by strict automation logic that even the smallest deviation can break. Agents have a greater degree of autonomy, integration, and flexibility in execution that lets them competently handle repetitive tasks. Meanwhile, human operators concentrate on the creative and strategic parts of their roles.
For example, let’s say a customer sends a vague email request about a product they can’t remember or bother to search for. Without a precise keyword match, they might get a generic apology or no response at all. Conversely, an AI agent can use context clues to identify the requested product or suggest alternatives and provide shipping estimates. The customer gets entered as a potential lead in a CRM, or the agent drafts a personalized response if data on past purchases and interests is already available.
What AI Still Struggles With
There’s no denying that AI’s current capabilities are noteworthy. The speed of improvement and innovation in the field is also genuinely impressive. And yet, AI still has serious limitations and shortcomings it would be foolish to ignore. Here’s what overzealous users either overlook or brush aside.
Heavy data dependence
Training data defines AI models to a great extent. Any deficiencies in said data will be present and amplified in the AI’s decisions. Various problems arise, from inherent biases to decreased reliability as more real-world inputs diverge from training examples.
No concept of reality or truth
AI systems are probabilistic and eager to please. An LLM will never admit that it lacks the data to form a valid response. Instead, it will give an answer with a high statistical probability of being what the user wanted, whether it’s correct or feasible. An AI might invent patterns where there are none or recommend nonsensical, even dangerous solutions because it has no concept of reality.
Trouble understanding nuance
Give an LLM a text to analyze, and it will do a decent job at the surface level. However, it lacks the reasoning and experience to pick up on subtle cues, cultural norms, or ethical dilemmas. If you ever want proof that AIs struggle with higher quintessentially human concepts, treat one sarcastically and see it miss your barb completely. Or better yet, have it tell you a joke.
Inability to weigh and make complex decisions
AIs’ decision-making processes are goal-oriented. Give an agent a desired outcome, and it will do everything in its power to achieve it. This falls apart when goals are uncertain or when the correctness of actions can be interpreted subjectively. Artificial intelligence can only propose solutions to complex problems and wait for human confirmation. It can’t decide which solution is the most appropriate when priorities and consequences are debatable.
Conclusion
Being aware of AI’s true capabilities and limitations is empowering. It lets ordinary users understand that the technology isn’t infallible and exercise their better judgment when dealing with questionable outputs. Meanwhile, professionals and decision-makers can leverage its strengths to boost productivity and adopt advantageous business strategies without taking AI’s suggestions at face value or having unrealistic expectations.
