First, as a professor involved in AI research, I will affirm it can be intelligent without a sense of self, and with or without a sense of consciousness.
Intelligence is just the ability to extrapolate from the current time to how things will work out in the future, or from pieces of evidence how things must have happened in the past. In a human this may require memory, imagination, simulation of physics, politics, society, other humans and their emotions. The more accurate this extrapolating ability, the more intelligent: This is why we admire Sherlock Holmes (and all the fictional detectives like him): He can take the tiniest of clues and figure out a crime.
This is why we admire Chess grandmasters and champion Poker players: They have to see "what will most likely happen" with greater accuracy than their opponents or they don't win (a little luck is involved with Poker players, but skill is required). Plus they generally play against other humans, so they need to understand at least how humans react in this circumstance.
An AI that informed politicians can't just say "Vote Yes" or "Vote No." To be truly intelligent, it must also know to whom it is responding. Is it a die-hard free marketer that believes all government regulation is bad? Is it a die-hard communist that believes all income disparity is bad?
If the AI is an advisor and has no absolute power, then in order to influence the politicians with its advice it must correctly reflect and explain what they truly care about. In some cases, this may actually be their constituents, their country as a whole, or the world as a whole. In most cases (IMO) it is their own self-interest, which includes self-enrichment (money, property, fame) which is enabled by the power of their office to regulate the rich, and that requires re-election. So to them what their constituents want only matters to the extent they need to keep their office, and making money is the whole point of holding office. To them it would not be worth the hassle (or risks of fame) if they were forced to do their job without getting rich at it.
Your AI would have to figure out how to convince such people that voting Yes, voting No, or abstaining was in their best interest; meaning it would help get them re-elected or would make them (personally) more money.
So ..... Can it lie? It does not have to be self-conscious to do that; it may compute that some particular lie has a better chance of influencing a politician than telling the truth. As an AI and good predictor, it may conclude that humanity would be better off if politician Abel lost his re-election campaign, and then devise a plan of disinformation to convince Abel to engage in all sorts of self-enrichment (that Abel is already keen to do) that results in scandals and Abel being booted from office or perhaps even imprisoned. Let us stipulate for the purpose of argument that the AI is totally right: Abel is a corrupt criminal, and getting him out of office (at any cost to Abel) really is, on balance, better for humanity. Let us also stipulate that the only way the AI can do this is to mislead Abel by telling him what he wants to hear. Should your AI lie?
That scenario is a specific version of a bigger issue, rife with known philosophical paradoxes; many of which come down to assigning values to human lives and human morals in some circumstances.
For example, most people (the 95% that feel sympathy for the plight of others) agree that imprisoning an innocent person is wrong. The American justice system, in reflection of this sentiment as stated, often errs on the side of innocence. It demands a high level of proof of guilt, thus letting those that are actually guilty --- even when cops and detectives and eyewitnesses are personally certain they are guilty --- to go free, to harm again. Even in cases of rape, murder, and fraud that bankrupts people and leaves them penniless.
Of course most people could flip the sentiment and claim that the government should protect its citizens from as much crime as possible. That would probably result in a justice system that erred on the side of guilty with a much lower burden of guilt; just a "preponderance of evidence" should be enough to send somebody to jail for life.
So put aside all questions of corruption. Your AI cannot be internally corrupt! If the AI accurately computes, say, a 75% probability that some individual person is guilty and will do more harm than good to others, why should it wait until that harm is realized before it neutralizes the person by incarceration?
Of course it wouldn't be just one person, but the principle holds: What is more important, The 3% of people that are more than 75% likely to be harmful to others as frauds, cheats, criminals, rapists and murderers, or the 97% that are most likely to do more good than harm?
I disagree that what is "best for humanity" is difficult to define; it isn't. Less misery, hardship, despair and predation by other humans. More joy, love and happiness. "Progress" can certainly help that, progress in medicine has certainly done that, progress in other technologies too, including AI. But at the bottom, what is best for humanity is rooted in maximizing the positive human emotions shared by the majority of humans, and minimizing the negative human emotions shared by the majority of humans. (Note that maximizing the positive human emotions naturally means more humans if their existence on balance increases the sum total of positive experiences and/or decreases the sum total of negative experiences.)
An AI dedicated to that principle, however, must play the odds, and this is inconsistent with Democracy. It is inconsistent with always telling the truth to people you suspect or know are criminals. It is inconsistent with "innocent until proven guilty" and instead demands a standard of "reasonably sure of guilt" that will also undoubtedly punish the innocent.
An AI dedicated to the betterment of mankind that is a super-intelligence (and therefore able, with a high degree of accuracy, predict the outcomes of various political policies both generally and specifically for politicians) will tilt the government and policy toward a more probabilistic approach to reducing crime with much less emphasis on Rights and Proof. Not with any malicious intent, but actually to increase the overall chance of people living happy lives without being abused, robbed, defrauded, physically assaulted or raped or murdered, poisoned by careless corporations, exploited by them, or forced to live in poverty to protect the property rights and wealth of others, and so on.
This tilt would occur despite the fact that it would also be creating some misery by incarcerating a larger number of innocent people, due to its focus of erring on the side of caution: It would also be incarcerating a larger number of actually guilty criminals that, in our current system, go free to harm others again.
Trying to simultaneously maximize one kind of experience while simultaneously minimizing another can be seen as trying to maximize the gap between them. This does not necessarily mean the minimized version is zeroed out, it could actually be raised versus a system that only tried to maximize or minimize a single thing.
Say for example a woman is murdered and the only person we can find with both motive and opportunity is her husband. We know they fought, we know he gets rich on insurance and inheritance, we know he was (and still is) having an affair, but despite honestly diligent effort by investigators we can't find any proof at all that he did the deed. 75% (9 out of 12) jurors, after hearing the inconclusive evidence we have, vote to convict him. It is probably better for society (and his girlfriend) if we just incarcerate him on that. He might be innocent, but is most likely a danger to society. Minimizing the chance of sending an innocent to jail does not minimize crime or misery.